Groq raises $650M after Nvidia talent deal

The headline reads like a contradiction: Groq raised $650M on June 22, 2026 to run inference datacenters seven months after Nvidia paid roughly $20B to license its core LPU IP and hire away founder/CEO Jonathan Ross, president Sunny Madra, and other engineers ^[1]. If the chip architecture and the people who designed it are gone, what exactly are investors buying?

The answer is that Groq is no longer selling a chip — it is selling capacity. The company already operates 13 data centers across North America, Europe, the Middle East and APAC, serving over five million developers and processing trillions of AI tokens each week ^[2]. That installed base, the customer relationships, and the physical infrastructure are the real asset that survived the Nvidia deal. The licensing agreement Nvidia signed was non-exclusive ^[3], so Groq retained the right to keep deploying and operating its technology even as Nvidia absorbed the IP and the founders.

What makes the financing unusual is that it is effectively guaranteed. Existing investors Disruptive and Infinitum committed to backstop the entire round if other existing investors decline their pro-rata shares — as one report put it, 'in some ways, the $650 million in funding is guaranteed' ^[4]. That structure tells you the raise is less a competitive market test of Groq's standalone value than a decision by its largest backers to keep the operating business alive.

The most striking irony of Groq's second act is hardware. The company that designed the Language Processing Unit to compete with Nvidia GPUs is now building its inference cloud on Nvidia silicon. Groq's new LPX system is a liquid-cooled appliance that houses Nvidia's LPU 3 chips, and the $650M will fund fitting out existing data centers with that hardware ^[5]. The firm that absorbed Groq's IP is now also a hardware vendor to Groq's pivot.

The technical stakes here are about memory architecture, which is where developer YouTube concentrated its attention. The dominant framing in those breakdowns is SRAM versus HBM: Groq's original LPU leaned on large on-chip SRAM for low-latency, deterministic inference, while Nvidia's GPUs rely on high-bandwidth memory (HBM) that can become a bottleneck for latency-sensitive serving. Those creator analyses read the Nvidia move less as a simple licensing transaction and more as a capability transfer — vertical integration across memory, inference, and frontier talent — which is also how the broader trade press has framed Groq's reorientation toward an inference 'neocloud' business ^[6]. The LPU 3 spec circulating in coverage — 92 lanes at 112 Gbps per lane (2.5 Tbps bidirectional) and 500 MB of onboard SRAM, debuted in March 2026 — shows the SRAM-heavy design philosophy lives on, just now inside Nvidia's product line ^[5].

For Groq, dependence on a supplier that is also the entity that hollowed out its team is the defining execution risk of the pivot.

The deal's structure is now a policy test case. In March 2026, Senators Elizabeth Warren and Richard Blumenthal opened an inquiry arguing the Nvidia-Groq arrangement is the latest example of 'reverse acquihiring' — a strategy that, in their words, 'involves acquiring control of a company's key assets without acquiring the company itself' ^[7]. The concern is that by licensing the IP and hiring the leadership rather than buying the company outright, Nvidia may have sidestepped the premerger notification that a conventional ~$20B acquisition would trigger.

Not everyone reads it as anticompetitive overreach. Industry analysts argue the move is strategically rational: by absorbing Groq's technology and talent, Nvidia secures its ecosystem and widens its 'competitive moat' in AI inference, keeping the next wave of inference innovation inside its own stack and out of reach of rivals ^[8]. Both readings agree on the underlying mechanic — Nvidia bought the parts of Groq that mattered to it — and disagree only on whether that should be legal without review. Either way, the inquiry could set precedent for how Big Tech consolidates AI talent and IP without ever filing a merger.

The most charged conversation about this deal is not about megawatts — it is about who collects. The center of gravity for that anxiety is secondary-market investors, where the dominant question is how the roughly $20B from the Nvidia deal is split and whether common shareholders see any of it. Holders worry the distribution favors taxes, retained staff, and preferred investors, leaving early common holders effectively wiped out, while contrarian voices counter that vested holders receive pro-rata payouts and that the $20B is small relative to projected AI infrastructure spend. There is even debate over whether the remaining 'Groq 2.0' cloud shell carries real value, with reports of the cloud business being explored for sale at around $1B.

The new $650M round sharpens that tension rather than resolving it. Because Disruptive and Infinitum are backstopping the entire raise ^[4], fresh capital flows in on terms set by the largest existing investors — and any holder who can't or won't follow their pro-rata risks dilution as 'Groq 2.0' is recapitalized. The funding trajectory frames the stakes: Groq raised $640M at a $2.8B valuation in August 2024, then $750M at a $6.9B post-money valuation in September 2025, and has now raised over $3B in total ^[9]. Tellingly, the valuation on the new $650M round was not disclosed ^[1]— an omission that, against two prior up-rounds, reads loudly to the investors trying to figure out what their shares are now worth.

Groq's three growth rounds tell the story of a company whose capital kept climbing even as its strategic position inverted. In August 2024 it raised $640M at a $2.8B valuation; in September 2025 it raised $750M at a $6.9B post-money valuation, led by Disruptive with BlackRock, Neuberger Berman, DTCP, Samsung, Cisco and Infinitum participating; and in June 2026 it raised $650M with no valuation disclosed ^[9]. Across these rounds the company has now raised over $3B total ^[9].

The inflection point sits between the second and third rounds: Nvidia's ~$20B December 2025 deal — its largest on record — that took the IP and the founders ^[10]. The earlier rounds funded a chip company climbing toward a $6.9B mark; the latest funds an inference-cloud operator whose valuation its backers chose not to publish ^[1]. Read alongside the operating targets — 200 MW of capacity by end of 2027, up from 13 existing data centers ^[2]— the trajectory shows capital being redeployed from designing silicon to renting it.