Samsung 2nm Yields Hit 70%: Qualcomm Talks, TSMC's Record Ramp & AMD Helios

Jamie Cole (0:00)

AI Hardware & Chips Daily News I'm Jamie Cole, thanks for joining me. Today the Samsung comeback, Ken Tuan MGAA challenged TSMC's foundry dominance.

Samsung's 2nm yields have crossed 70%.

That's the number that changes the foundry conversation. And it's the signal worth paying attention to right now. For most of early 2026, Samsung's 2nm Gator round process was sitting below 50% yields. That's a foundry node that's commercially marginal. 70% is different. It's not TSMC territory. TSMC is running 80-90% on 2nm, but it's close enough that serious customers are starting to run the numbers. The key implication is this. Samsung has just reopened a competition that most of the industry had quietly written off. The clearest evidence of that reopening is Qualcomm. The company is in serious talks with Samsung about a dual foundry setup for the Snapdragon Gen 6, splitting production between TSMC and Samsung's 2nm node. That's not speculation anymore. There have been direct executive level meetings in South Korea. The important distinction here is between hedging and genuine competition. Qualcomm exploring a second source tells you two things. One, TSMC capacity is tightening fast enough that a major customer is worried about allocation. Two, Samsung yields have improved enough that Qualcomm's engineering teams believe the quality risk is manageable. The unresolved piece is pricing. Samsung isn't likely to discount below 10 to 15% on a new node, given recent yield progress. No agreement is confirmed, and a pricing impasse could still store the deal.

TSMC isn't watching passively. The company is targeting 110,000 wafers per month at 2nm by mid-2027, up from 90,000 at the end of 2026

That's a 22% capacity jump. 3nm is heading to 210,000 wafers per month at the same time. Here's the thing. To put that in context, this is the largest foundry capacity ramp ever attempted in a single cycle. TSMC is allocating 70-80% of a 60-64 billion dollar capital expenditure budget to advanced nodes. The risk that comes with that scale is real.

110,002 nm wafers per month assumes AI demand stays strong and customer deployment timelines hold.

If packaging becomes a choke point, which is a live concern, utilization could slip on the logic side even as wafer output grows.

On the demand side, AMD's Helios platform is generating the kind of customer signals that validate the capacity race. OpenAI, Meta and Anthropic have all committed to multi-gigawatt deployments. AMD is projecting data center revenue to more than double year-on-year in 2027 AMD claims Helios delivers 30% more tokens per dollar than competing platforms. That metric matters because the inference economics conversation is shifting the whole industry's framing. Raw peak performance is no longer sufficient. Total cost of inference is what hyperscalers are optimizing for now. Consider this.

The caveat is that the 30% figure hasn't been validated independently.

Execution on yield and power delivery at hyperscale is still an open proof point. Away from the TSMC-Samsung duel, there's a subplot worth tracking. Euclyd, a Dutch inference chip startup, has closed a 200 million euro Series A with Samsung as co-lead investor. Samsung is also manufacturing Euclyd's Craftworks Silicon in South Korea, targeting first customer deliveries in 2027

The signal here is geopolitical as much as technical. European AI sovereignty has produced a lot of policy documents and very few real chips. Euclyd is an attempt to change that, backed by Sirius Capital and a credible manufacturing partner. No first Silicon has shipped yet, and no independent benchmarks exist. Whether this becomes the next ASML, or remains a well-funded aspiration, is a question 2027 will start to answer. You see, the through line across all of this is a foundry market in genuine structural transition. Samsung's yield improvement is real, but incomplete. Qualcomm's interest is real, but not yet a signed contract. TSMC's capacity expansion is the most concrete development, but it carries over capacity risk if AI deployment timelines slip. The metrics to watch are specific. Can Samsung close the yield gap to TSMC's 80-90% range by mid-2027? Does the Qualcomm Samsung deal move from torts to a confirmed wafer allocation? And can the back-end, power modules, advanced packaging, co-os keep pace with the logic node ramp? Those three proof points will tell you whether the Samsung come back is structural or just a useful negotiating lever for customers who need TSMC to worry about competition again. Thanks for listening. This podcast was built using AI technology, a YesWe production.