A Product Nine Months in the Making
Nvidia's Groq 3 LPX interactive inference accelerator entered full production on August 24, 2026, announced at the Hot Chips conference. The chip is purpose built for agentic AI workloads, the fast, high volume token generation that chatbots and autonomous agents depend on, and slots into Nvidia's Vera Rubin platform alongside the company's Vera CPUs and Rubin GPUs.
The name is not a coincidence. In December 2025, Nvidia paid Groq roughly 20 billion dollars to license its chip architecture and hired away founder and CEO Jonathan Ross along with key engineers, a deal structured as licensing plus hiring rather than a formal acquisition. Nine months later, hardware built on that licensed design is shipping under Nvidia's own name.
The Numbers
Nvidia's own benchmarks position Groq 3 LPX as a speed play first, aimed at the workloads where response latency, not raw model size, decides whether an agentic AI product feels usable.
| Metric | Groq 3 LPX figure |
|---|---|
| Output speed | 3,400 tokens per second (Gemma 4 31B benchmark) |
| Context window | 100,000 tokens |
| Rack density | 256 LP30 chips per rack |
Nebius is the first AI cloud provider to adopt the chip, with racks going live later this year alongside Nvidia's Vera CPUs and Rubin GPUs on its Token Factory inference platform. SpaceX has also signed on as an early Vera Rubin platform customer for its own AI infrastructure.
The Same Chip, a Different Owner
A week before this launch, Groq itself raised 350 million dollars at a 3.5 billion dollar valuation, half what it commanded at its September 2025 peak, in a round that included Nvidia as an investor. The company has repositioned from a chip manufacturer into an AI inference cloud operator, competing on running other companies' models rather than selling its own silicon, a pivot forced by losing the founder and team that built its hardware roadmap.
Groq 3 LPX is the other half of that same story. The technology and the people Nvidia licensed and hired in December are now the product Nvidia sells under its own brand, to customers Nvidia signs directly. What started as two separate companies with two separate roadmaps has become one roadmap, owned by the incumbent the challenger was built to beat.
What This Confirms for a Buyer
Any procurement strategy built around a second inference vendor specifically to reduce Nvidia dependence needs to treat this launch as evidence, not theory. A licensing-plus-hiring deal that draws none of the merger scrutiny a formal acquisition would trigger can still end with the challenger's core technology shipping as an incumbent product nine months later.
The practical takeaway is to weight vendor diversification decisions by how defensible a challenger's independence actually is, not just by its current market position. A chip startup with a founder-led team and no incumbent investor on its cap table is a different risk than one where the dominant player already holds equity, licensed IP, or a board seat, regardless of how the two compare on paper today.
Read next: Nvidia's Land Grab Signals the Real AI Bottleneck | Ohio Ratepayers Won't Pay for OpenAI's 8GW Campus
