Hot Chips 2026: Nvidia presents Groq 3 LPX architecture and unveils its first third-party inference benchmark — LP30-based rack already in production, company says
Nvidia presented the Groq 3 LPX architecture at Hot Chips 2026, showcasing its first third-party benchmark. The LP30-based rack, already in production, achieved 3,431 tokens per second on a 100K-context Gemma 4 31B workload. Nvidia acquired Groq's technology for $20 billion in 2025, integrating it into its product lineup.
How this was made
The 30-second read
Why it matters
The announcement provides the first independent performance validation of Nvidia's new inference accelerator, potentially influencing AI hardware buying decisions.
Market read
New AI inference hardware could shift competitive dynamics and affect Nvidia's valuation.
What to watch
Supply‑chain constraints for SRAM chips and lack of HBM could limit adoption in data‑center scale.
Background
Nvidia acquired Groq in Dec 2025 for $20 billion, integrating its LP30 chip into the new LPX rack presented at Hot Chips 2026.
Ticker impact
Nvidia unveiled the Groq 3 LPX rack and released the first third‑party benchmark showing 3,431 tokens/sec, four times faster than the next‑fastest endpoint.
potential short‑term upside as investors price in the performance advantage
First‑hand benchmark data and production status are fresh, material information for a major AI hardware player.
Market effects
Strengthens Nvidia's position in the AI inference hardware market, may pressure GPU rivals.
U.S. tech sector could see modest gains; limited immediate effect on broader markets.
Highlights continued AI hardware race, relevant to global AI infrastructure investors.
Counterpoint
Benchmark numbers may not translate to real‑world workloads; competitors could catch up quickly.
Key entities
- CompanyNvidia
US‑listed AI hardware and GPU leader.
- CompanyGroq
Acquired AI inference chip designer.
