According to monitoring by 1M AI News, Jonathan Ross, founder and former CEO of AI inference chip company Groq, recounted the behind-the-scenes process of Nvidia's $20 billion deal. He initially entered Jensen Huang's office to procure about 100,000 Blackwell GPU chips. He pitched an idea to Huang: AI data centers should not use the same hardware for all tasks; training, like bulk freight, requires '18-wheel trucks' (GPUs), while inference is better suited for flexible and efficient 'vans' (Groq's LPU chips). Three days later, Huang proactively called back, not to discuss the order, but to promote deeper collaboration. On Christmas Eve, three weeks later, the deal was finalized. Nvidia obtained licensing for Groq's chip technology, and Ross along with other core employees joined Nvidia, with Ross taking on the role of Chief Software Architect. At the time of the deal, Groq's annual revenue was just under $100 million, with only $3 million in revenue for 2023 and a loss of $88 million. At the GTC 2026 conference, Huang announced the Groq 3 LPU inference chip based on Groq technology and the Groq 3 LPX rack containing 256 LPUs, which has entered full production and is expected to start deliveries this summer. Ross stated, 'This is on the scale of Nvidia, not a pilot.' This marks Nvidia's shift from a single GPU dominance to a full-stack 'AI factory' solution encompassing CPUs and LPUs. The deal has also drawn the attention of the U.S. government. Democratic Senators Elizabeth Warren and Richard Blumenthal questioned whether the agreement was 'structured to evade antitrust scrutiny' and jointly wrote to Huang requesting more disclosure on the terms. Nvidia responded that it did not acquire Groq, but only purchased a non-exclusive license for the intellectual property and hired talent.