Nvidia Ramps Up Samsung Foundry Orders for LPU Chips

Nvidia's $20 billion deal for Groq was less a corporate acquisition and more a strategic absorption of talent and intellectual property. The transaction was structured as a non-exclusive licensing agreement for Groq's LPU technology and a mass talent transfer, which included founder Jonathan Ross and a significant portion of the engineering team, effectively neutralizing a key competitor in the low-latency inference market. Groq's Language Processing Unit (LPU) is architected fundamentally differently from a GPU. It utilizes a single-core, deterministic design with large amounts of on-chip SRAM instead of external HBM, a design choice that eliminates the unpredictability of traditional caches and schedulers to deliver extremely low-latency performance for sequential AI inference tasks. The increased orders are a strategic move for Nvidia to diversify its supply chain beyond TSMC, which currently handles the majority of its GPU production. This move mitigates risks associated with single-foundry dependence, especially with the ongoing capacity constraints on advanced packaging technologies like TSMC's CoWoS, which is crucial for high-performance AI accelerators. Samsung Foundry's role is critical, especially as reports indicate its 4nm process yield for logic dies has surpassed 90%. This improvement in manufacturing maturity makes it a viable and attractive option for high-volume production of advanced chips like the LPU, providing Nvidia with a credible high-performance alternative to TSMC for specific product lines. This foundry diversification is a key trend among fabless semiconductor giants. Intense demand for AI accelerators is causing a power shift in the foundry ecosystem, with companies like Nvidia commanding a larger share of cutting-edge wafer production, even surpassing mobile chip orders. Having multiple high-volume foundry partners like Samsung and potentially Intel in the future provides greater negotiation leverage and supply security. The comparison to Intel's potential production capacity on its 18A-class nodes highlights the long-term competitive landscape. Intel is aggressively marketing its foundry services and has started high-volume manufacturing on its 18A process, positioning itself as another key player in the high-performance computing manufacturing space that major designers like Nvidia are evaluating for future products. For the Silicon Valley talent market, Nvidia's "acquihire" of the Groq team demonstrates the intense competition for specialized hardware engineering expertise. The growth in AI is putting immense pressure on the industry to innovate faster, leading to a projected shortfall of design workers by 2030 and making the retention of experienced teams that can execute on novel architectures a top strategic priority.