Nvidia booked $14.4 billion of its $17 billion Groq license as goodwill — a figure the plaintiffs in a new Delaware lawsuit say proves the deal's real value was the people and technology the Groq board handed over, not cash returned to shareholders.

Two former Groq engineers, Benjamin Serebrin and Joshua Rubin, filed a proposed class action in Delaware's Court of Chancery against Groq's former board and former CEO Jonathan Ross. The suit, unsealed Monday, alleges the board structured a $20 billion transaction — a $17 billion license fee plus a separate $3 billion Nvidia stock pool for engineers who transferred — in a way that kept billions from shareholders who did not move to Nvidia.

Groq told CNBC the suit is "meritless" and that its licensing agreement with Nvidia "delivered exceptional value for Groq, our investors, and our employees."

What do Nvidia's own filings reveal?

Nvidia's annual report (10-K) breaks the deal down starkly:

• $14.4 billion booked as goodwill, "primarily attributable to the workforce and future development of the licensed technology"
• $2.5 billion for the technology itself
• No customer contracts, existing products, or equity interests purchased
• $13.0 billion paid at closing, with $4 billion (inclusive of imputed interest) due within one year

The goodwill figure is roughly 85% of the $16.9 billion Nvidia booked — about 5.8x the value assigned to the licensed IP. Nvidia paid far more for the LPU design team and its future roadmap than for the silicon design itself.

The technology in question is concrete. The LP30 packs 500MB of SRAM per die with 150 TB/s of bandwidth, is fabbed on Samsung's 4nm process, and is rated by Nvidia at up to 1.23 FP8 PFLOPS. A full LPX rack holds 256 LPUs — 128GB of SRAM, 40 PB/s of bandwidth, 315 FP8 PFLOPS aggregate — and serves as a decode co-processor for the Vera Rubin NVL72. The LP30 matches the next-gen LPU Groq roadmapped in August 2023 on Samsung's SF4X node with its first-generation Tensor Streaming architecture.

At Hot Chips in August, Igor Arsovski — Groq's former chief architect, now Nvidia VP of hardware — called the moment "a pinch me moment for the Groq team that's now integrated into the Nvidia group," adding the rack was already in production.

Why does the structure matter legally?

The suit alleges Ross and other top employees could "take a discount on those shares and be paid separately for following the technology to Nvidia," and that the board denied some shareholders a vote. Nvidia hired "nearly all" of Groq's engineers — as many as 200. The complaint also argues the $17 billion license was taxable as income at Groq, and that a subsequent $350 million Series A at a $3.5 billion valuation effectively repriced the residual company above what bought-out shareholders received. Four "conflicted funds" on the board — BlackRock, Social Capital, Infinitum, and Disruptive, not named as defendants — allegedly profited by staying on after the buyout.

The plaintiffs concede that "no Delaware decision has directly answered the question this case raises." The outcome could shape future license-plus-hire deals, including Nvidia's pending Hugging Face acquisition, which similarly combines a purchase price with equity for employees who join.

What are the regulators doing?

The New York Times reported in September that the DOJ sent Nvidia a formal request for information, with a fine possible. Nvidia called the Groq story "a prime example of the American system working as designed." Senators labeled it a "reverse acqui-hire" in February, and FTC Chair Andrew Ferguson said in January the agency is "beginning to examine these acqui-hires to make sure they are not an attempt to get around" merger review.

Meanwhile, Groq continues operating. It runs 13 data centers, expects to scale from 54 to 200+ MW in 2027, joined Nvidia's Cloud Partner program in August, and plans to deploy Groq 3 LPX racks with Dell — built on the very technology it licensed away. Nvidia names Nebius as the first AI cloud to adopt LPX, with racks online later this year, and its roadmap already lists the NVFP4-capable LP35 and the LP40 for Feynman alongside Rubin Ultra.