Foundry Market Up 29% — And TSMC Took Almost All of It

A bigger pie, sliced more unevenly

The global foundry business — the contract manufacturing of chips designed by other companies — grew 29% year over year in the second quarter. In semiconductors, that kind of expansion is rare. What matters more is how it was distributed: TSMC came in at roughly 73% market share, while Samsung Foundry remained stuck around 7%. The market is growing, and the gap is growing faster.

The driver is no mystery. Demand for AI accelerators and high-performance computing silicon has pulled the industry's biggest designers — Nvidia, AMD, Broadcom, Apple — toward 3nm and 5nm-class leading-edge nodes. Only a handful of fabs can produce those nodes at volume with predictable yield, so the orders concentrated. In this cycle, yield at the leading edge is market share, more bluntly than at any point in the past decade.

Why Samsung is stuck at "the 7% wall"

Korean media have started calling it exactly that — a wall at 7%. The reasons go well beyond raw process technology.

Switching costs compound. Foundry is a multi-year commitment. A fabless customer builds its process IP, design kits and verification flows around one manufacturer, and moving to another means redoing that work. If a customer had a bad yield experience in one generation, it rarely gambles again in the next. What TSMC accumulated over decades isn't only equipment; it's the cost of leaving.

Samsung competes with its own customers. Samsung makes memory, mobile application processors and finished phones. For a potential client, handing over a chip design can feel like handing it to a rival business unit. TSMC's founding promise — that it never competes with its customers — was never just branding; it was the core sales argument, and it still is.

Capacity is being locked up in advance. Reporting around the latest numbers describes AI chip buyers paying whatever it takes to reserve leading-edge capacity ahead of time. Once lines are booked out by pre-orders, a challenger doesn't even get a chance to bid. And in a supply-constrained market, undercutting on price stops being a viable weapon — customers are optimizing for certainty of supply, not cost per wafer.

How this differs from Samsung's memory playbook

The contrast with DRAM is instructive. In memory, Samsung has held the top position for years by controlling technology, cost and capacity simultaneously. But memory is a standardized commodity produced at scale, while foundry is closer to a bespoke service business. Wins depend on design-enablement teams, ecosystem partners, IP library breadth and hand-holding through tape-out. A manufacturing mindset struggles against a service business — a lesson the last ten years have repeatedly delivered.

Intel's foundry push offers a parallel worth noting. Despite enormous capital spending and government support, Intel has also found it slow going to land marquee external customers. Two of the world's most capable chipmakers are demonstrating the same thing at once: money and process nodes alone do not buy foundry share.

What actually changes for buyers

For anyone purchasing chips, concentration is bad news. When one supplier controls three-quarters of the market, pricing power sits with that supplier. Leading-edge wafer price increases tend to work their way through AI servers and GPUs and, eventually, into consumer devices. Smaller fabless startups with modest volumes risk falling to the back of the allocation queue, which can push product launch schedules out by quarters.

There's also a geopolitical dimension. The more advanced logic production clusters in one region and one company, the more a single earthquake, export restriction or geopolitical shock could reset AI infrastructure timelines worldwide. That risk is precisely why governments across the US, Europe, Japan and Korea keep subsidizing domestic fab capacity — the concentration numbers make the policy case for them.

Where a realistic opening might be

For Samsung, chasing TSMC across every node is probably not the winning strategy. Selective differentiation looks more plausible: stabilizing yield on its 2nm gate-all-around process, building out advanced packaging, and offering combinations TSMC structurally cannot — most notably bundling high-bandwidth memory and logic together in a turnkey deal. As AI chips become as memory-bound as they are compute-bound, the ability to co-design and co-supply HBM alongside logic may become genuinely valuable rather than a talking point. Whether that translates into share is unproven, but it is a differentiated angle rather than a slower copy of the leader.

The takeaway

A 29% growth figure reads like an industry-wide boom, but it mostly documents how unevenly the boom is being shared. The battleground in foundry has shifted away from nanometer marketing and toward a harder triangle: yield, capacity and trust. The metric to watch next quarter isn't a decimal point move in Samsung's share — it's whether a major AI customer publicly commits volume to a second source.