At GTC 2026, Jensen Huang gave Samsung Foundry a public thank you. He didn’t mention TSMC.
After six years of relentless effort, the alchemist's first success gleams in the dawn light.
“Samsung Electronics is manufacturing the Groq 3 Language Processing Unit (LPU) chip for us. We are truly grateful to Samsung,” Huang said from the stage. The chips ship in Q3 2026. The name that dominates every advanced-node conversation was absent.
This is not a one-off design win. It is a structural shift. Samsung Electronics' 4nm process yield has surpassed 80%, the threshold industry insiders recognize as the standard for a mature process, according to reporting from Gate. That yield figure has reached parity with TSMC, per SamMobile. The immediate result is a string of orders from Groq, IBM, and Baidu. The long-term result is the dismantling of TSMC’s unilateral pricing power on mature advanced nodes at the exact moment hyperscalers need a second source.
The discovery of a second viable route shifts the balance of power from the old monopoly.
The six-year crawl to parity
Samsung’s 4nm process entered its sixth year of mass production in 2026. The early years were a mess. Defect densities drove customers to TSMC and forced a leadership crisis inside Samsung’s foundry division. That period is over. Sedaily reports that process variables have decreased and defect patterns are thoroughly controlled. RC Delay, the phenomenon that slows data movement inside chips, has been reduced by approximately 26% from previous levels through the addition of specialized stacks.
The 4nm node now offers a full platform with multiple threshold voltage options, from uLVT for high-speed performance chips to HVT for low-power designs that prevent leakage current. This is not a single-purpose process tweaked to land one customer. It is a complete platform ready for volume production across multiple design points.
How Samsung turned a liability into a hedge
The yield improvement did not happen by accident. Samsung made a deliberate trade: it delayed its 1.4nm process mass-production schedule from 2027 to 2029 to concentrate resources on optimizing 2nm and improving yields at existing nodes, Gate reports. The industry consensus framed this as a retreat. It was triage. Samsung traded leadership glamour for high-volume, high-margin pragmatism.
The synergy play adds a structural advantage. Samsung’s foundry division now produces HBM4 base dies, locking in a logic-plus-memory package that TSMC cannot replicate with the same integration. Sedaily reports industry projections that HBM4-related revenue could exceed 30 trillion won, roughly $22 billion. Samsung also became the first company in the world to mass-produce and ship sixth-generation HBM4, a milestone confirmed in March 2026. A customer buying 4nm logic from Samsung can source HBM4 from the same supply chain. That bundling power is a durable advantage on mature advanced nodes.
The end of unilateral pricing power
The consensus is framing this as “Samsung is back,” implying a head-to-head performance war with TSMC on 2nm and below. That analysis misreads the strategy. Samsung’s victory condition is not beating TSMC on transistor density. It is becoming the industrial default for “good enough” advanced silicon. TSMC wins the science fair. Samsung wants to own the factory floor.
Here is what’s confirmed: Groq entrusted all production of its LPU from the first to the third generation to Samsung’s 4nm process, Sedaily reports. IBM and Baidu have placed 4nm orders with Samsung Foundry, according to Gate. Here is what this means: 4nm is the new mature advanced node, good enough for the vast majority of AI inference and non-leading-edge compute. With a credible second source, the hyperscalers now have the leverage to demand dual-sourcing for supply-chain sovereignty.
That structural shift will cap TSMC’s pricing power on mature nodes for the first time in a decade. The prediction here is specific: within 18 months, Samsung will secure at least two of the so-called Magnificent Seven as 4nm logic customers, not for leading-edge performance but for supply-chain resilience. This dual-sourcing reality will compress TSMC’s gross margins by 200 to 300 basis points on these nodes. The Groq win is the proof-of-concept. IBM and Baidu are the validation.
The operator’s takeaway
For chip designers, the calculation has changed. A viable second source for 4nm-class nodes exists now, with multiple Vth options and a complete design platform.
For investors, TSMC’s margin narrative faces a new headwind. Samsung’s foundry business is expected to turn profitable no later than the second half of 2026, according to Gate’s market reporting.
For the industry, the AI supply chain is actively de-risking from a single point of failure. The era of “TSMC or nothing” for mature advanced nodes is over.
Jensen Huang did not mention TSMC because he didn’t need to. The story is no longer about who owns the angstroms. It’s about who can provide sovereign, reliable supply at scale. Samsung just answered that question.
