Nvidia’s $20B Groq Deal Heralds The Era Of ASICs—Has The Commoditization Of AI Chips Begun?
Nvidia’s $20B Groq deal validates the need for ASICs in the era of inference. While Nvidia's GPUs are protected by a wide technological moat, ASICs risk rapid commoditization, implying Nvidia’s historically high margins might face compression.
This article represents the author’s opinion only, is not financial advice, and is intended for entertainment purposes only. The author holds no beneficial position in any of the companies mentioned, receives no compensation for writing this article, and has no business relationship with any of the companies mentioned.
In March 2024, Nvidia CEO Jensen Huang told an audience at Stanford University with characteristic confidence that his company’s chips were "so good that even when the competitor's chips are free, it's not cheap enough." He was arguing that Nvidia's GPUs are so efficient that they face no real competition when considering the total cost of running an AI data center. Indeed, the company’s skyrocketing valuation in recent years has been fueled by the conviction that its integrated hardware and software ecosystem offers value no rival can match.
Less than two years later, however, that very narrative has been challenged by Nvidia’s own capital allocation. The company’s recent decision to pay approximately $20 billion to non-exclusively license technology from Groq and hire its founders marks a significant strategic pivot. By spending billions to access a competing architecture, Nvidia is effectively acknowledging that its general-purpose GPUs may no longer be the most economically viable solution for the rapidly expanding inference market.
Nvidia structured the transaction as a non-exclusive licensing agreement combined with the hiring of key personnel, including founders Jonathan Ross and Sunny Madra. This mirrors the "acqui-hire" playbook employed by Microsoft and Amazon in recent years to absorb top-tier talent without triggering the regulatory scrutiny of a formal acquisition. However, spending roughly three times Groq’s recent $6.9 billion valuation indicates that Nvidia is willing to pay a substantial premium to integrate Groq’s specific approach immediately. If Nvidia’s internal roadmap for inference silicon were fully addressing its customers' shift toward efficiency, there would be little justification for such a deal.
This move validates a bifurcation in the semiconductor market that Nvidia had previously been reluctant to acknowledge. For years, the company has downplayed the threat of Application-Specific Integrated Circuits (ASICs), insisting that the rapid evolution of AI models requires the flexibility that only programmable GPUs can provide. Huang has frequently argued that custom chip projects are often abandoned because they simply cannot keep pace with Nvidia’s innovation cycle.
The Groq deal reinforces the opposing view: that as AI models move from training to mass deployment, the market prioritizes cost and speed over flexibility. Groq’s ASICs, named Language Processing Units (LPUs), use a deterministic architecture that strips away the scheduling overhead of a GPU, offering a level of throughput and latency that general-purpose hardware struggles to match. By bringing this technology in-house, Nvidia is effectively admitting that ASICs are required for cost-effective inference. This undermines its previous argument that the GPU is the universal solution for all AI computing.
This pivot is underscored by the shifting behavior of Nvidia’s largest customers. Hyperscalers have been aggressively developing their own ASICs to reduce reliance on Nvidia’s high-margin hardware. This dynamic came into focus on November 24, when The Information reported that Meta is considering a multi-billion dollar deal to buy Google’s Tensor Processing Units (TPUs). If formalized, this partnership would mark a significant turning point. Meta has historically been one of the most aggressive purchasers of Nvidia’s GPUs; if a customer of that magnitude is exploring Google’s custom silicon for inference workloads, it signals the first cracks in Nvidia’s monopoly.
The Groq transaction closely follows Nvidia’s move to hire the leadership of networking startup Enfabrica for $900 million. Both deals aim to shore up specific vulnerabilities in the AI stack that competitors like Broadcom are actively exploiting. Broadcom’s custom silicon business has quietly become a dominant force, empowering hyperscalers to design the very chips that displace Nvidia’s hardware. By absorbing Groq’s intellectual property, Nvidia attempts to offer an internal alternative to these custom solutions, aiming to retain customer spend that might otherwise bleed to ASIC providers.
While these deals may help solidify Nvidia’s position as the undisputed leader in AI infrastructure, they introduce new questions about the durability of its moats and margins in the era of inference. The company has historically commanded gross margins near 75 percent for data center chips because its GPUs effectively had no substitutes in the training market. However, inference is a more price-sensitive market where 'good enough' performance at a lower price point is often the winning strategy.
If Nvidia is forced to compete with increasingly commoditized ASICs, its margins will likely face compression. The company’s valuation has been built on the assumption that it can maintain monopoly-like pricing power for years to come. The Groq deal suggests that Nvidia’s leadership sees a future where that is no longer guaranteed. Ultimately, this license agreement tacitly admits that rival chips have caught up, proving that competitors like Groq once dismissed as 'not cheap enough' even when free are now worth substantial sums to bring in-house.
Follow me on X for frequent updates (@chaotropy).
