When the biggest shovel-maker buys the guy building a different kind of shovel, regulators start asking whose goldmine this really serves.
The Summary
- The DOJ is investigating Nvidia's $20 billion licensing deal with AI chip startup Groq, questioning whether the structure was designed to dodge antitrust scrutiny
- Nvidia controls roughly 80% of the AI accelerator market; Groq builds chips optimized for inference (running models) rather than training them
- The probe signals escalating regulatory attention on how dominant AI infrastructure players consolidate power through strategic partnerships
The Signal
Nvidia didn't buy Groq outright. Instead, it structured a licensing agreement worth $20 billion that keeps Groq technically independent while giving Nvidia deep access to its inference chip technology. The DOJ wants to know if this was creative dealmaking or regulatory arbitrage.
The distinction matters. An acquisition would trigger mandatory Hart-Scott-Rodino antitrust review. A licensing deal, depending on how it's structured, might not. Nvidia has already faced scrutiny over its CUDA moat and its relationships with cloud providers. This probe suggests regulators think the company might be using licensing agreements as a backdoor to acquisitions.
"Nvidia controls the picks and shovels for training AI models. Now regulators are asking if they're locking up the tools for running them too."
Groq represents a different bet than Nvidia's core business. While Nvidia's H100s and upcoming Blackwell chips dominate model training, Groq's Language Processing Units are optimized for inference: the actual work of running trained models at scale. Inference is where the revenue is long-term. Every ChatGPT query, every agent action, every real-time AI application runs on inference chips.
If Nvidia can control both the training infrastructure and the inference layer through partnerships that don't trigger full regulatory review, it cements end-to-end dominance in the AI compute stack. That's the kind of vertical integration that makes antitrust lawyers reach for their keyboards.
The timing is notable. This comes as:
- Inference workloads are growing faster than training workloads across hyperscalers
- Groq was positioning itself as an Nvidia alternative, not an Nvidia partner
- The EU and UK are already investigating Nvidia's partnerships with cloud providers and AI companies
The Implication
Watch for a wave of regulatory scrutiny on "partnership" deals in AI infrastructure. If licensing agreements become the new acquisition strategy, expect regulators to adapt their frameworks. For companies building AI infrastructure, the message is clear: strategic partnerships with market-dominant players will get second looks, regardless of how you structure the paperwork.
For everyone betting on a competitive AI chip market, this matters. If Nvidia can neutralize would-be competitors through licensing deals that fly under regulatory radar, the inference chip market may consolidate before it ever really opens up.