FirstMoveTUE 25 AUG 2026
TODAYWORLDMARKETSAIAI SKILLSSPORT
First Move  ·  AI Hardware & Infra  · 
The build-out — Tuesday morning, 25 August

Nvidia Groq 3 LPX Inference System Enters Production, Promises 35x Throughput Per Megawatt

The AI compute build-out sees semiconductor revenue surging to $1.6 trillion in 2026, though rising memory costs and power infrastructure bottlenecks continue to challenge rapid expansion.

EDITION 2026-08-25 · EVERY CLAIM SOURCED · GROUNDED VIA SEARCH AT PUBLISH
ShareWhatsAppXLinkedIn

The story

This week, Nvidia announced its Groq 3 LPX inference system has entered full production, with deployments at cloud provider Nebius slated for later this year. The system, produced by Samsung Foundry on a 4nm process, integrates 256 language processing units and aims to accelerate AI inference workloads.

Benchmarks by Artificial Analysis indicate the LPX system can process 3,400 tokens per second using Google's Gemma 4 31B model, a speed Nvidia claims is four times faster than the nearest alternative platform. When paired with Nvidia's Vera Rubin processors, the combined system is estimated to deliver up to 35 times the throughput per megawatt.

This launch comes as Gartner forecasts global semiconductor revenue to reach $1.6 trillion in 2026, with AI data centers driving a significant portion of this growth. However, the industry is also grappling with escalating memory costs and the immense power requirements for these high-density AI clusters.

Silicon

Groq 3 LPX

Maker: Nvidia (produced by Samsung Foundry)

What: AI inference system with 256 language processing units, processes 3,400 tokens/second with Google Gemma 4 31B model, 4x faster than alternatives.

For Whom: Nebius and other cloud providers deploying AI inference workloads.

The build-out

Data center power infrastructure investment

Who: Nvidia, Cloverleaf Infrastructure

Scale: Hundreds of millions of dollars investment, securing over 10 GW of planned data center power capacity

Where: Various locations for large-scale data center projects

Supply & policy signals

Nvidia raising AI server prices by over 15% for Grace Blackwell and Vera Rubin systems

Implication: Reflects surging memory costs, with server DRAM doubling in Q1 2026, making memory a quarter of high-end AI server rack costs.

Gartner forecasts DRAM revenue to increase 246.6% and NAND flash revenue 371.9% in 2026

Implication: Indicates strong demand and sharply stronger pricing, suggesting more dollars are spent without necessarily proportional growth in deployed capacity.

Taiwan experiencing an AI-driven power crunch, impacting foundry capacity and chip prices

Implication: Foundry capacity is tightening, and chip prices are climbing into the second half of 2026, despite a nearly 90% year-over-year increase in Taiwan's July ICT export orders led by AI servers.

TSMC increased its 2026 capital expenditure forecast to $60-64 billion

Implication: Highlights ongoing massive investment to meet long-term capacity plans and technology development roadmap driven by strong demand for AI applications.

What we'll be watching

Reporting + analyst voices: grounded via Google Search at publish time.