Nvidia's reported megadeal with Groq isn't a traditional acquisition, but rather a carefully structured licensing agreement that delivers tax advantages alongside inference technology and engineering talent. The company licensed Groq's language processing unit (LPU) technology and hired key employees, but explicitly did not purchase Groq's equity, existing products, or customer contracts. This separation allows Nvidia to claim tax-deductible goodwill tied to the acquired workforce and anticipated future technology development, effectively reducing the transaction's effective long-term burden.

Groq publicly announced the agreement on December 24, 2025, calling it a non-exclusive license covering Groq's inference technology. The brief announcement said founder Jonathan Ross, president Sunny Madra, and other employees would join Nvidia, while Groq would remain independent and continue operating its GroqCloud service without interruption. However, neither company disclosed the deal's financial terms or identified every transferred asset at that time.

When Nvidia filed its annual report, the company revealed more details about how it structured the transaction. The filing classified most of the recorded value as goodwill, an accounting asset representing value that cannot be neatly assigned to identifiable property like a patent or existing product. Nvidia said this goodwill primarily reflected the acquired workforce and anticipated future development of the licensed technology.

What Makes This Deal Structure Different from a Standard Acquisition?

A traditional acquisition typically brings an operating business under one owner, including contracts, revenue, employees, products, intellectual property, and corporate control. Nvidia instead separated the pieces strategically. The company secured technology rights and recruited the team positioned to advance that technology, while Groq kept its corporate identity, cloud operation, and customer-facing business.

This distinction matters for competition and regulation. The transaction presents itself as non-exclusive licensing, yet it still transfers many of the capabilities that made Groq a serious technical challenger in AI inference. Regulators and lawmakers must now decide whether that separation preserves meaningful competition or only its legal outline.

The filing added a crucial detail: the goodwill is tax deductible. This means Nvidia expects the relevant tax basis to generate deductions under applicable tax rules. A deduction is not an immediate refund, and it does not make the transaction free. Instead, it generally reduces taxable income over the period permitted by tax law, subject to Nvidia's income and tax position.

How Does Groq's Technology Fit Into Nvidia's Broader Inference Strategy?

Groq designed its LPU around predictable execution for inference workloads, the stage when a trained AI model responds to prompts, generates tokens, or makes predictions. This differs from training, which emphasizes enormous parallel computations across large clusters. Interactive inference places greater weight on response latency, token generation speed, memory movement, and consistent performance.

Groq's architecture relies heavily on software scheduling and on-chip SRAM, a fast memory type located close to processing resources. That design can reduce delays caused by repeatedly moving data between processors and external memory. However, Groq's architecture also creates scaling challenges because large models can require many chips to hold and execute the workload.

Nvidia's existing systems offer complementary resources Groq lacked. GPUs can handle broad parallel workloads, while Nvidia's networking, software, and data center platform connect large numbers of accelerators. The companies began testing a disaggregated approach before the agreement, assigning different phases of an inference request to hardware optimized for each task.

"The teams experimented with running different portions of a workload on Nvidia GPUs and Groq LPUs," said Jonathan Ross.

Jonathan Ross, Founder of Groq

That technical path became visible at GTC 2026, when Nvidia introduced Groq technology as part of the Vera Rubin platform rather than treating it as a disconnected accelerator. The Groq 3 architecture divides inference into prefill and decode stages. Prefill processes the initial prompt, while decode generates the response tokens that users see. Nvidia can assign each stage to the resources best suited for it.

Steps to Understanding How Nvidia Integrates Groq's Technology

• Licensing the Full Software Stack: Nvidia licensed not just Groq's LPU hardware architecture, but also Groq's compiler and full software stack, which carries knowledge about splitting models across many LPUs and scheduling their execution.
• Recruiting the Engineering Team: Groq engineers joined Nvidia's Dynamo team, which develops software for orchestrating inference across data center infrastructure, ensuring the technology integrates seamlessly with Nvidia's existing platform.
• Connecting Three Critical Layers: The arrangement connects chips, networking, and orchestration software, the three layers that determine practical performance, since a faster accelerator alone cannot deliver its theoretical advantage if the surrounding system leaves it waiting for data.

Nvidia executive Ian Buck explained that the company licensed Groq's full software stack, not just the hardware design. Groq engineers also joined Nvidia's Dynamo team, which develops software for orchestrating inference across data center infrastructure. This mechanism explains why Nvidia wanted more than a passive patent license; the compiler and engineers carry crucial knowledge about splitting models across many LPUs and scheduling their execution.

Nvidia already controls a widely used software environment through CUDA and owns networking technology gained through Mellanox. The company also sells integrated data center systems instead of isolated processors. Groq's design can now become another component inside that platform, allowing customers to gain a specialized inference option without adopting an entirely separate hardware and software environment.

The tax-deductible goodwill structure reveals how Nvidia approached the deal beyond pure technology acquisition. By classifying the transaction this way, Nvidia obtained not only a new inference architecture and experienced engineers but also an accounting asset whose tax treatment can lower the transaction's effective long-term burden. This advantage arose from the same structure now attracting competition questions from regulators and lawmakers evaluating whether the separation truly preserves meaningful competition in the AI inference market.

" }