NVIDIA Didn't Buy Groq. It Converted Groq Into a Customer.

On Christmas Eve 2025, NVIDIA announced its largest deal in company history. It was announced on Christmas Eve deliberately.

CNBC's headline called it a "$20 billion acquisition." The official press release — Groq's, not NVIDIA's — called it a "non-exclusive technology licensing agreement." The distinction matters enormously, and it was designed to be confusing.

NVIDIA did not buy Groq. NVIDIA licensed Groq's LPU chip architecture and GroqWare compiler stack, acqui-hired ~90% of the engineering team, brought founder and CEO Jonathan Ross in as NVIDIA's Chief Software Architect, and left the Groq brand and GroqCloud business intact as a legally independent entity. The deal value was $20.6 billion in cash — the figure comes from Disruptive CEO Alex Davis; Groq's official press release omits the number. This is NVIDIA's largest completed deal by approximately 3x (Mellanox: $6.9B in 2020).

What NVIDIA actually bought.

Groq's LPU — Language Processing Unit — was the most technically credible non-GPU architecture for transformer inference. The differentiation was specific: up to 230MB of on-chip SRAM with approximately 80 TB/s on-die bandwidth, all execution compiler-scheduled at compile time, no external HBM round-trips. The result was approximately 5.8x faster than H100 on per-request inference latency and approximately 3.6x cheaper per million tokens. These are large numbers.

The architecture was built on a single insight: GPU inference is bottlenecked by memory bandwidth. Every time you need weights or activations, you're waiting for HBM transfers. The LPU eliminates the wait. NVIDIA's GPU architecture is optimized for training and mixed workloads. The LPU was optimized for inference only — not a training replacement, but genuinely superior at the thing that most enterprise AI deployments actually spend money on.

The GroqWare compiler is the part that makes this work commercially. Static scheduling at compile time — no runtime decisions, fully deterministic execution — is what produces the latency guarantees. The chip architecture matters, but the compiler is the moat.

NVIDIA licensed both. Then it hired the people who wrote the compiler.

The non-exclusive license as competitive theater.

The "non-exclusive" framing means Groq can theoretically license the same LPU IP to AMD, Intel, or any other chipmaker. This is the structure NVIDIA presented to regulators and senators as evidence of preserved competition.

The practical reality, which multiple analysts described in nearly identical language: "buying a Formula 1 car without the engine or pit crew." The engineers who built the LPU and maintain the GroqWare compiler are now at NVIDIA. A bare patent license without the people who implement it is not a usable competitive weapon. The competitors who might license the IP — AMD, Intel, any newcomer — would have to rebuild from scratch the compilation toolchain that Groq spent eight years developing. NVIDIA, by contrast, now has both the IP and the team.

Jonathan Ross.

Ross built Google's original Tensor Processing Unit project before leaving to found Groq in 2016. The TPU architecture went on to become Google DeepMind's primary training infrastructure for the Gemini series. Groq's LPU went on to become the most discussed GPU alternative in enterprise inference. Ross is now NVIDIA's Chief Software Architect.

Jensen Huang now has, under one roof, the architects of GPU inference dominance (NVIDIA's original team), the architect of the leading TPU lineage (via training infrastructure heritage), and the architect of the only commercially deployed LPU. Every significant inference architecture story of the past decade has one of its originators at NVIDIA.

Separate reports indicated OpenAI was attempting to recruit Ross away from NVIDIA after the deal closed. Those reports are unconfirmed by primary sources.

The $650M raise.

In June 2026, six months after the deal, the reconstituted Groq — now led by new CEO Adam Winter, CFO Matt Eng, COO Alan Rice (formerly xAI and Meta Datacenters), and an incoming CTO and CPO — raised $650 million led by Disruptive and Infinitum.

The use of proceeds: outfit 13 existing data centers with NVIDIA LPX hardware, scale toward 200 MW by 2027.

The company that was NVIDIA's most credible inference alternative raised $650 million specifically to become NVIDIA's large-format hardware customer. The conversion is complete. Groq is a well-capitalized proof-of-concept that NVIDIA's LPX platform works at enterprise inference scale. It is not competition. It is reference architecture with a brand.

The circular structure is real: Disruptive invested $500M+ in Groq since founding, became the deal's largest beneficiary (~3x step-up on last-round valuation, ~7x on Series D), and is now Groq's chairman and lead investor in the $650M raise. Disruptive's incentive is for Groq to scale as fast as possible — which means buying NVIDIA hardware as fast as possible.

The regulatory question the deal was structured to avoid.

Under the Hart-Scott-Rodino Act, companies must file premerger notification and observe a waiting period before completing an acquisition above specified thresholds. The trigger is an acquisition of voting stock, assets, or non-exclusive rights. The legal analysis is contested, but NVIDIA structured the Groq deal as a technology licensing agreement rather than an asset acquisition — which likely sidesteps HSR notification requirements.

On March 19, 2026, Senators Elizabeth Warren, Richard Blumenthal, and Ron Wyden sent a letter to Jensen Huang questioning whether the deal "is an attempt to avoid antitrust laws" and set an April 3 response deadline. The FTC announced in January 2026 it would examine "merger in disguise" arrangements across the technology sector — citing Microsoft's acqui-hire of Mustafa Suleyman and Inflection AI staff as the precedent case. No formal DOJ or FTC investigation specifically targeting NVIDIA's Groq deal has been opened as of available reporting.

The Microsoft-Inflection deal was approximately $650M. NVIDIA-Groq is $20.6B for ~90% of an engineering organization plus the primary licensing asset. If the HSR structure holds without regulatory challenge, the licensing-instead-of-acquiring approach becomes the playbook for AI consolidation going forward — one in which the antitrust review that would normally accompany an acquisition of this scale simply never triggers.

The timing.

The Christmas Eve timing drew near-universal analyst characterization as deliberate news-burial — a holiday when markets are closed and press coverage is minimal. The framing has not been challenged. No NVIDIA or Groq spokesperson offered an alternative explanation.

The deal closed on the quietest business day of the year. It was the largest deal NVIDIA has ever done. The company that announces every product at a major keynote chose not to issue a press conference for its $20.6B transaction and instead let Groq post a press release on groq.com while most of the technology press was opening presents.