Groq: The Complete Guide

NVIDIA now sells a rack called the NVIDIA Groq 3 LPX, in full production from August 2026; "Groq and LPU are used under license from Groq, Inc." 8. Groq itself became an NVIDIA Cloud Partner and plans to be among the first to bring those racks to market 9 10. In August 2026 Groq raised $350 million at a $3.5 billion valuation, about half its September 2025 value 3 11.

Funding: $640 million at a $2.8 billion valuation (August 2024, led by BlackRock) 14; $750 million at $6.9 billion (September 2025, led by Disruptive) 11; $650 million in June 2026 12; and $350 million called a "Series A" at $3.5 billion in August 2026, with planned participation from NVIDIA 3.

The 2023 Samsung deal was for a 4 nm chip at Samsung's new Taylor, Texas plant, which Groq said would scale "from 85,000 to over 600,000 chips" 15. No Groq-branded 4 nm chip has shipped; the chip that emerged is NVIDIA's Groq 3, which press reports say Samsung makes on 4 nm 19 5.

NVIDIA's Groq 3 chip has 500 MB of SRAM and 150 TB/s of memory bandwidth; each LPX rack holds 256 of them, 128 GB of SRAM plus 12 TB of DDR5 memory 22. Groq quotes 315 petaflops of FP8 compute and 1,000 tokens a second per user per rack 23.

The first chip measures about 725 mm² with about 28.6 billion transistors, according to press reports citing analysts 24. An analyst estimate puts its wafer cost at "likely less than $6,000", well below leading-edge chips 26. At NVIDIA's GTC in March 2026, press reports say, Jensen Huang said Samsung "manufactures the Groq 3 LPU chip" 5.

Bell Canada plans 500 MW across six sites, with Groq as its exclusive inference provider 18. The Saudi site moved from "Dammam" in the 2025 release to "Riyadh" on the current list; Groq has not explained the change 16 23.

The catch is memory size: 230 MB per first-generation chip 20. An analyst counted 576 chips, in 8 racks of 9 servers with 8 chips each, to serve one Mixtral model 26. Groq has not published an official chip count for a given model size.

Press reports give Groq 3 at 98 billion transistors on Samsung's SF4X process 5; NVIDIA's page gives 500 MB of SRAM where one press report says 512 MB, so we use NVIDIA's figure 22 19. The GroqCard has up to nine chip-to-chip connectors 20.

Groq does not publish hardware prices, and neither NVIDIA's LPX page nor the GroqCard datasheet gives one 22 20. Groq's price list changes often; we show only models whose prices appear on the official page.

Groq's remaining advantage is its cloud: 13 data centres, millions of developers and exclusive deals such as Bell Canada 3 18. But it now runs NVIDIA hardware too, so its edge is increasingly the service rather than the chip 9.

Groq has not published revenue in any release we read 12. Press reports, citing The Information, said in July 2025 that Groq cut its 2025 revenue forecast from about $2 billion to about $500 million after delays to the Saudi deal 34.

11. No price for the NVIDIA deal appears in either company's release 4 8.
12. The US Justice Department is reported to be examining the deal; no official statement yet 35.
13. Groq's own speeds and outside tests differ: 500 against about 477 tokens a second for GPT OSS 120B 6 30.
14. Yann LeCun was named a technical adviser in 2024 14.
15. The Saudi site name changed from Dammam to Riyadh 16 23.

Groq's dependence on NVIDIA is growing: it is certified to "design, deploy, and operate NVIDIA accelerated computing" and plans NVIDIA Groq 3 LPX and Vera Rubin racks in its data centres 9 10. The 2025 revenue forecast cut reported by the press shows how much one customer region mattered 34.

For India's own chip plants and plans, see our India's semiconductor push guide.

The open questions are regulatory, the reported US review of the NVIDIA deal 35, and commercial, whether GroqCloud can grow against far larger clouds 3.