NVIDIA's Groq 3 LPX rack is now in full production and will be deployed alongside the company's Vera CPUs and Rubin GPUs at neocloud provider Nebius later this year, Nvidia senior director Dion Harris told reporters, according to CNBC.

The rack represents the commercialization of technology from Nvidia's $20 billion purchase of Groq's assets in December, its largest deal on record. The deal is part of a broader race to speed up AI inference—the serving of trained models to users—which has become a major battleground as cloud providers compete to deliver faster, more responsive AI applications. Each liquid-cooled rack packages 256 Groq chips. Nvidia said the rack can deliver 3,400 tokens per second, citing a benchmark from Artificial Analysis. Groq's chips are manufactured by Samsung, unlike Nvidia's own GPUs, which are made by Taiwan Semiconductor Manufacturing Co.—a detail that also spreads Nvidia's silicon supply across two foundries rather than one.

Bull Case

The $20 billion Groq deal represents a manageable financial risk for NVIDIA Corporation (NASDAQ: NVDA) relative to the strategic capabilities it could add to the company. Bernstein analyst Stacy Rasgon told CNBC that Nvidia is financially strong enough to absorb a deal of this size with little impact on its overall position. That gives Nvidia the flexibility to make large strategic investments while expanding into new AI chip technologies without materially straining its balance sheet.

Groq's technology extends Nvidia's dominance across the full AI compute stack rather than creating a rival product line. LPX handles low-latency inference while Nvidia's GPUs continue handling training and large-context processing. The rack can pair with Vera Rubin chips without customers changing their CUDA workflows, deepening the software lock-in that has made Nvidia hard to displace.

Execution has been fast, which matters in a market moving this quickly. Nvidia went from announcing the Groq purchase in December to full production and a named customer, Nebius, in eight months. That pace shows Nvidia can absorb acquired technology and ship it rather than let it stall in integration.

Nvidia's inference speed now has a quantified edge over a key rival's approach. The Groq rack's 3,400 tokens per second compares with the 750 tokens per second OpenAI has promised for its Cerebras-powered "Ultrafast" mode. It gives Nvidia a benchmark it can point to as cloud providers shop for faster, more responsive AI inference.

Bear Case

NVIDIA paid $20 billion for Groq's technology and talent without acquiring the company itself. Since Nvidia licensed Groq's technology and hired its employees rather than purchasing Groq outright, investors have less visibility into the full value of what Nvidia obtained and whether the transaction can generate returns proportionate to its unusually large price tag.

Nvidia is running many of these bets simultaneously, so it cannot give full focus or enough cash to any single one. Beyond Groq, Nvidia has committed up to $100 billion to OpenAI and $5 billion to Intel. It has also put smaller sums into Crusoe, Cohere, and CoreWeave, plus a similar $900 million licensing deal for Enfabrica's team, a broad spread of bets where not everyone will pay off.

LPX is a narrow, specialized product rather than a broad platform. The chip's design makes it well suited to inference but brings limitations that make it less so for other tasks like training. Its addressable use case is therefore inherently smaller than Nvidia's core GPU business.

The rollout still rests on one named customer, Nebius. It is the only cloud provider confirmed to deploy the Groq rack so far, and until more commit, it is unclear whether the technology sees broad adoption or remains a niche addition.

Hedge Fund Data

Insider Monkey's database shows NVIDIA was held by 285 hedge funds in Q2 of 2026, up from 275 in the first quarter. Rival AMD was held by 164 hedge funds.

Conclusion

This is NVIDIA extending its reach into AI inference at a cost too small to move its own financial results, which is exactly why it matters strategically rather than financially. Optimists believe Nvidia will win by moving fast, outperforming competitors' technology, and keeping customers tied to its CUDA software. Pessimists fear it will struggle by spreading money across too many projects at once, buying technology instead of whole businesses, and relying on a product only one customer uses.