On Aug. 25, Samsung Electronics said its foundry division has begun full-scale production of the Groq 3 LPX, an inference chip Nvidia designed for AI workloads, a contract win that South Korean media said pushes the foundry business to the edge of profitability. The announcement marks the first time Samsung’s advanced process has carried a major Nvidia product in the current AI cycle, and it gives the company a foothold in the market where it has trailed Taiwan’s TSMC for a decade.
The chip, based on Nvidia’s architecture and tailored for reasoning and inference, is manufactured on Samsung’s leading-edge process node. Samsung did not disclose volumes, but people familiar with the arrangement said the initial production run is substantial, filling capacity that had been idle while the foundry division waited for anchor customers. The ramp is expected to improve utilization at Samsung’s fabrication plants in Hwaseong and Pyeongtaek, which have operated below breakeven for several quarters.
Samsung’s foundry business has been the weak spot in a company otherwise defined by memory chips. It lost the advanced-process race to TSMC during the transition to 7-nanometer manufacturing, and its attempts to catch up with aggressive technology claims were undermined by yield problems that forced it to push back production schedules. Customers drifted to TSMC, and Samsung’s foundry share fell from a peak of roughly 20 percent to the single digits for cutting-edge work.
The Groq 3 LPX order changes the narrative in an important way. It is a volume product from the world’s largest AI chip designer, produced on a node that Samsung has now brought to manufacturing maturity. Korean media reported that the yield rate has stabilized at commercially acceptable levels, and that Nvidia has qualified additional products for Samsung production. The company’s own executives have said the foundry division is “one step from turning a profit,” a statement backed by the improving capacity picture.
The win did not come easily. Samsung spent years courting Nvidia, which had resisted moving advanced products to a supplier whose record was uneven. The breakthrough came from a combination of price, capacity, and a specific manufacturing advantage: Samsung’s ability to offer a process tuned for the power-efficiency profile that inference chips need. Nvidia, facing a capacity squeeze across the industry, was willing to give Samsung a second look.
The deal also reflects a structural shift in the AI chip market. Training chips, Nvidia’s traditional stronghold, are built in enormous volumes on the most advanced nodes available. Inference chips, the fast-growing segment of the market as AI moves into production, are smaller, more numerous, and more sensitive to cost and power. That profile suits a foundry with competitive pricing and available capacity, even one whose technology trails the leader.
Industry analysts cautioned that one product does not restore Samsung’s position. TSMC still commands more than half of the foundry market and controls the packaging technology, CoWoS, that the most advanced AI chips require. Samsung has invested in its own advanced packaging, but its share of that critical step remains small. The Groq 3 LPX order, analysts said, gives Samsung a proof point to show other customers, not a guarantee of a broader shift.
For Nvidia, the arrangement has a strategic dimension. The company has long pursued a multi-supplier strategy to avoid depending on any single foundry, and it has pushed both TSMC and Samsung to expand capacity. Orders like the Groq 3 LPX give Nvidia negotiating strength and insurance against disruptions. The chip is part of Nvidia’s push into the inference market, where it faces competition from custom silicon built by its own customers.
Samsung’s management has signaled that foundry is central to its plan to move beyond the memory cycle that dominates its earnings. The company is spending tens of billions on new fabs in Texas and South Korea, banking on a future where logic chips and advanced packaging generate a growing share of revenue. The Groq order is the first visible return on that bet.
The stakes extend beyond Samsung and Nvidia. A credible second foundry for advanced AI chips would reshape pricing and supply dynamics across the industry, giving chip designers alternatives to TSMC and giving governments worried about geographic concentration a more balanced map of production. The U.S. and South Korea have both subsidized advanced manufacturing in recent years, and the Groq 3 LPX ramp demonstrates that the capacity they funded can attract real orders.
For now, Samsung’s next goal is profitability. Executives have said the foundry division should reach breakeven as utilization improves, and the Groq production run accelerates that timeline. Whether it tips into consistent profit depends on winning more customers like Nvidia, and those conversations, according to people familiar with Samsung’s plans, are already underway.
