Git Informed
Git
Informed
All newsBriefing
Editorial front page
FinalAI-edited source brief

Groq raises $650M to reinvent itself as an inference neocloud after licensing its chip tech to Nvidia

After a $20B deal that sent Groq's founders to Nvidia and paid out shareholders, the startup is rebuilding around inference hosting — on Nvidia's own silicon.

Published 8 sources0 Reddit7 web88% confidence

What matters

  • Groq raised $650M from existing investors (Disruptive, Infinitum) to fund its 'Groq 2.0' inference neocloud pivot after licensing its LPU technology to Nvidia for ~$20B in December 2025.
  • Nvidia has already productized the licensed IP as the Nvidia Grok LPU 3 inference processor, shipping in a liquid-cooled LPQ appliance with 32 trays since March 2026.
  • The Nvidia deal sent Groq's founding CEO Jonathan Ross and senior leaders to Nvidia and paid shareholders ~$7.6B (~$64/share); Groq remains an independent entity.
  • Groq's inference cloud is now shifting to run on Nvidia GPUs rather than its own LPU silicon, with interim CEO Adam Winter and interim CFO Matt Eng leading the transition.
  • Antitrust scrutiny from Senators Warren and Blumenthal remains unresolved; Nvidia's rapid productization of the LPU 3 could complicate any regulatory remedy.

What happened

Groq confirmed on June 22, 2026 that it has raised $650 million from existing investors — led by Disruptive and Infinitum — to fund its pivot into an AI inference neocloud. The raise closes a chaotic chapter: in December 2025, Groq struck a roughly $20 billion licensing deal with Nvidia for its Language Processing Unit (LPU) technology, a transaction that sent CEO Jonathan Ross and other senior leaders to Nvidia and paid shareholders approximately $7.6 billion in cash (about $64 per share). (One outlet, MarketScreener, reported the licensing figure at $17 billion.)

Nvidia wasted no time productizing the licensed IP. In March 2026, the chip giant debuted the Nvidia Grok LPU 3, an inference processor that ships as part of a rack-size, liquid-cooled appliance called the LPQ. Each LPQ contains 32 trays, each hosting three LPU 3 units alongside a CPU and networking gear. The LPU 3 includes automatic clock-drift correction and 92 data lanes running at 112 gigabits per second — concrete evidence that Groq's architecture has already reached the market under Nvidia's brand.

Meanwhile, Groq is rebuilding as what it internally calls "Groq 2.0": an inference neocloud that hosts open-weight models like Llama, Mixtral, Qwen, and DeepSeek for developers and enterprises. The company is now led by interim CEO Adam Winter and interim CFO Matt Eng, with no permanent CEO named. Backers Disruptive and Infinitum backstopped the full $650M round, effectively guaranteeing its close.

Why it matters

Groq's pivot is one of the more unusual second acts in recent AI history. The company went from designing chips that competed with Nvidia to running a services business on top of Nvidia's platform — all while its core intellectual property now belongs to its largest competitor. The bet is that inference workloads (the processing that happens after an AI prompt) are growing fast enough to support a vertically focused neocloud even without proprietary hardware differentiation.

The deal also raises unresolved antitrust questions. Senators Elizabeth Warren and Richard Blumenthal have scrutinized the Nvidia–Groq transaction, and that inquiry remains open. Nvidia's rapid productization of the LPU 3 adds urgency: the licensed technology is already shipping in commercial hardware, which could complicate any regulatory remedy.

For Groq's existing inference-cloud customers, the transition from LPU to Nvidia GPU infrastructure may affect performance characteristics, pricing, and capacity. The company's fate now hinges on customer retention and whether inference demand grows fast enough to sustain a neocloud without unique silicon.

Public reaction

No Reddit or community discussion threads were captured for this story at the time of writing. Public reaction is limited to coverage in tech and business press, with no measurable grassroots developer sentiment available.

What to watch

  • Whether Groq's existing inference-cloud customers stay now that the underlying hardware is shifting from LPU to Nvidia GPUs.
  • How regulators resolve the Warren–Blumenthal antitrust inquiry, especially given that Nvidia has already shipped the LPU 3.
  • Whether Groq can compete against established neoclouds like CoreWeave and hyperscaler inference offerings without proprietary hardware.
  • Who Groq names as permanent CEO and whether the interim team can execute the pivot.

Sources

Public reaction

No Reddit or community discussion threads were captured for this story at the time of writing. Public reaction is limited to coverage in tech and business press, with no measurable grassroots developer sentiment available.

Open questions

  • Will Groq's existing inference-cloud customers stay now that the underlying hardware is Nvidia GPUs rather than Groq's LPU?
  • How will regulators resolve the Warren–Blumenthal antitrust inquiry into the Nvidia–Groq deal, given that the LPU 3 is already shipping?
  • Can Groq compete as a neocloud against established players like CoreWeave and hyperscaler offerings without proprietary hardware differentiation?

What to do next

Developers

Evaluate Groq's inference cloud API for hosting open-weight models like Llama, Mixtral, Qwen, and DeepSeek, and benchmark latency and cost against alternatives such as Together AI and Fireworks AI.

Groq's value proposition has shifted from LPU speed to neocloud convenience; developers should verify whether the new Nvidia-backed stack still offers competitive inference performance.

Founders

Study the 'not-acqui-hire' structure as a potential exit path that licenses IP and transfers talent while preserving the corporate entity for a pivot.

The Groq–Nvidia deal demonstrates a novel transaction type that returns capital to investors while leaving room for a second act — relevant for hardware startups facing platform consolidation.

PMs

Assess inference-cloud vendor risk by mapping which providers depend on a single chip supplier and how that affects pricing, availability, and roadmap commitments.

Groq's pivot to running on Nvidia silicon creates a dependency that PMs should factor into vendor selection for inference workloads.

Investors

Scrutinize the $650M raise terms and Groq's path to profitability as a neocloud without proprietary hardware, and monitor the unresolved antitrust inquiry.

The investment thesis has fundamentally changed from chip differentiation to services scale; the open regulatory question adds downside risk.

Operators

If currently using Groq for inference, validate service continuity, SLA terms, and hardware transition timelines with Groq's account team.

The shift from LPU to Nvidia GPU infrastructure may affect performance characteristics, pricing, and capacity availability for existing customers.

How to test

  1. 1Sign up for Groq Cloud or obtain API credentials.
  2. 2Deploy an open-weight model (e.g., Llama 3) via Groq's inference endpoint.
  3. 3Run a standardized set of prompts and measure token-generation latency, time-to-first-token, and cost per million tokens.
  4. 4Repeat the same benchmark on a comparable inference provider (Together AI, Fireworks AI, or direct Nvidia GPU hosting) for comparison.
  5. 5Check whether Groq discloses which hardware (LPU vs Nvidia GPU) underpins your specific endpoint.

Caveats

  • Groq may be mid-transition from LPU to Nvidia GPU infrastructure; performance characteristics could change without notice.
  • Pricing and capacity availability may shift as the $650M raise is deployed.
  • The antitrust inquiry could affect the licensing agreement terms and Groq's long-term access to its own technology.