DOJ가 AI 추론 칩 전문 기업인 Groq에 대한 Nvidia의 200억 달러 인수를 정밀 검토하고 있으며, 이는 거래 완료에 규제 위험을 야기하고 있다.

Nvidia spent $20 billion late last year to license Groq's AI accelerator technology and recruit key members of its engineering team in an expansive deal structured to operate beneath regulatory scrutiny. The acquihire left Groq's inference-as-a-service business technically intact, but the US Department of Justice launched an antitrust probe into the arrangement this week, according to reports from The New York Times.

Without its core engineering staff, Groq functions largely as Nvidia's subsidiary. Yet Nvidia maintains the transaction exemplifies American innovation policy. "The Groq story is a prime example of the American system working as designed to promote innovation, reward entrepreneurs, and benefit consumers," Nvidia said in a statement. "The law is designed to encourage America's startup ecosystem and promote the fundamental rights of inventors and workers to pursue their dreams."

Nvidia acquired two principal assets: Groq's mature silicon and the talent required to advance it. Groq built its reputation on SRAM-heavy dataflow accelerators that accelerated large language model inference to hundreds and now thousands of tokens per second—a capability Nvidia's GPU-based systems had struggled to match independently. While faster than GPUs, Groq's accelerators faced throughput limitations. The performance relationship resembles an F1 car (Groq's LPUs) against a city bus (Nvidia GPUs); combined, they approximate a sport pickup truck.

At GTC in March, Nvidia unveiled LPX racks powered by 256 Groq-3 accelerators, lightly modified versions of Groq's existing Groq-2 designs—a realistic timeline given three months to production. CEO Jensen Huang promised the Groq-3 combined with Vera Rubin GPU racks would optimize inference across all workload types. However, disaggregated compute architectures are not unique to Groq. Cerebras is building comparable systems with AWS and AMD, SambaNova partners with Intel, and d-Matrix combines its in-memory compute platform with Nvidia GPUs in similar configurations.

Whether the deal genuinely harmed competition remains contested. Several US senators have argued that transactions engineered to evade regulatory review warrant scrutiny anyway. Proving material competitive harm may prove difficult given the fragmented landscape. Even if the DOJ litigated successfully and unwound the deal—hardly unprecedented for Nvidia—the outcome might prove inconsequential.

Before announcing the Groq licensing deal, Nvidia was already establishing an ecosystem centered on networking. In late 2024, the company contributed its MGX rack designs to the Open Compute Project, allowing any chipmaker to deploy their processors in Nvidia GPU-designed racks. In mid-2025, GPUzilla opened its high-speed interconnect technology through NVLink Fusion licensing, enabling any chipmaker to slot designs into Nvidia's racks. A regulatory unwinding would cost Nvidia direct control of LPU development and licensing revenues, but the Groq LPX racks would persist. Groq would simply join Nvidia's growing partner ecosystem. In fact, the acquihire ensures that even if regulators succeed, alternative suppliers remain positioned to fill any gap.