NVIDIA and Groq explained: why Jensen Huang compares the licensing deal to Mellanox

NVIDIA’s Groq deal is not a classic acquisition. In December 2025 NVIDIA licensed Groq’s low-latency inference technology on a non-exclusive basis and hired its leaders, paying $13 billion at closing and owing $4 billion more within a year. On February 25, 2026, Jensen Huang said NVIDIA would add Groq to its platform as an accelerator, as it once added Mellanox networking.

What did NVIDIA get from Groq?

A license and people, not the company. NVIDIA received a non-exclusive license to Groq’s chip technology and hired much of its leadership; Groq kept its own business.

In its press release of December 24, 2025, Groq said that founder Jonathan Ross, president Sunny Madra and other team members would join NVIDIA to help develop and scale the licensed technology. Groq continues as an independent company with Simon Edwards as chief executive, and its GroqCloud service keeps running. Because the license is non-exclusive, Groq can still license the same technology to others.

The companies did not announce a price. CNBC reported a cash deal of about $20 billion. NVIDIA’s annual report (10-K) for fiscal 2026 gives its own figures: $13.0 billion paid at closing and $4 billion, including imputed interest, payable within one year. NVIDIA says it acquired no customer contracts, existing products or equity in Groq, yet it recorded $14.4 billion of goodwill and a $2.5 billion developed-technology asset with a five-year life. For comparison, in September 2025 Groq had raised $750 million at a post-money valuation of $6.9 billion in a round led by Disruptive.

What exactly did Jensen Huang say about Mellanox?

He made the comparison twice on the call for the fourth quarter of fiscal 2026 and left the details for NVIDIA’s GTC conference in March.

In his prepared remarks, according to the call transcript, Huang said: “As we did with Mellanox, we will extend NVIDIA’s architecture with Groq’s innovations to enable new levels of AI infrastructure performance and value.” Answering an analyst who expected Groq to add a decode-specific product, he was more concrete: “What we’ll do is we’ll extend our architecture with Groq as an accelerator in very much the ways that we extended NVIDIA’s architecture with Mellanox.”

On the same call he described the demand behind it. Huang said the industry had reached “the inflection of agentic AI” and argued that “in this new world of AI, compute equals revenues”: without computing capacity a provider cannot generate tokens, and without tokens it cannot grow revenue.

Why is Mellanox the benchmark?

Because Mellanox gave NVIDIA the networking it needed to sell complete data-center systems rather than chips alone. Huang is signalling that Groq should play a similar role for inference.

Mellanox, an Israeli maker of InfiniBand and Ethernet networking equipment, agreed to the deal in March 2019 at an enterprise value of about $6.9 billion; it closed in April 2020 at about $7 billion, after the US waiting period expired and regulators in the EU, Mexico and China approved it. NVIDIA now sells networking together with its GPUs. CFO Colette Kress said on the February call that networking revenue for fiscal 2026 exceeded $31 billion, more than ten times the level of fiscal 2021, the year NVIDIA acquired Mellanox.

Why does low-latency decode matter for AI inference?

Inference has two phases: prefill, when the model processes the prompt, and decode, when it generates the answer token by token. Decode speed sets how fast an AI agent can respond, and it is limited mostly by how quickly the chip can read model data from memory.

This is where Groq’s design differs from a GPU. According to Groq, its Language Processing Unit keeps memory and compute on the same chip: on-chip SRAM delivers more than 80 TB/s of bandwidth, while the external HBM used by GPUs runs at about 8 TB/s. The LPU is also deterministic: data flow is scheduled by the compiler in advance, so every execution step is predictable. The trade-off is capacity: on-chip SRAM holds far less data than HBM, so large models are spread across many chips.

NVIDIA had already worked on the other phase. In September 2025 it announced Rubin CPX, a GPU built for the context-processing (prefill) stage of long-context inference. Groq’s technology covers decode, the phase where agents wait for each next token.

How could NVIDIA integrate Groq’s LPUs?

As of mid-March 2026, before GTC, NVIDIA had not described the product, and two scenarios were discussed by analysts.

• A separate rack next to GPUs. Analysts at GF Securities, as reported by Wccftech, expected NVIDIA to show an “LPX” rack with 256 LPUs at GTC 2026. In that setup GPUs would handle prefill and LPUs decode.
• LPUs inside future GPUs. igor’sLAB described a more radical option: building LPU blocks into future architectures such as Feynman through hybrid bonding. It called this technically appealing but much harder in packaging, yield and thermal design, and considered the rack-scale route more realistic in the near term.

Both are expectations, not announcements. Huang himself only promised to share his “great ideas” at GTC.

What does the deal mean for investors?

For NVIDIA, the deal costs more than Mellanox, until then its largest purchase, and is a bet that inference will need specialized hardware alongside GPUs. For investors in private AI chip companies, it shows a new kind of exit: a license and hiring instead of a takeover.

Groq’s backers did not sell their shares, yet the deal worked for them like an exit. According to Axios, most Groq shareholders were to receive per-share distributions based on a $20 billion valuation, about 85% up front, about 10% in mid-2026 and the rest at the end of 2026. That valuation was roughly three times the $6.9 billion of the September 2025 round. How pre-IPO investors usually get liquidity is covered in how to exit a pre-IPO transaction, and why AI infrastructure draws so much venture money in why AI infrastructure has become one of the main venture topics.

• Product risk. In March 2026 there was no NVIDIA product with Groq technology yet; the Mellanox comparison is a statement of intent.
• Accounting. Most of the price sits on NVIDIA’s balance sheet as goodwill, which has to be written down if its value becomes impaired.
• Competition. The license is non-exclusive, so Groq can sell the same technology to others; what Groq did next as an independent company is covered in Groq 2.0.

Update (October 2026): At GTC in March 2026 NVIDIA presented the Groq 3 LPU and the Groq 3 LPX rack for its Vera Rubin platform, and on August 24 it said Groq 3 LPX was in full production, with Nebius as the first AI cloud to adopt it; each LPX rack holds 256 LPU accelerators with 500 MB of SRAM each.