Trendforce reported on May 27 that TSMC has informed customers to budget for up to a 15% price increase on 3nm process nodes in the second half of 2026, followed by another 5 to 10% jump in 2027. The list of customers eating the increase is short and familiar: Apple for the M-series and the iPhone, Nvidia for the Blackwell and Vera generations, AMD for the MI450 sampling pipeline, Broadcom for the custom ASIC business that is currently outgrowing Nvidia on percentage terms, and Qualcomm for whatever Qualcomm is doing this quarter.

Two facts are worth holding in mind together. The 3nm price hike is the third consecutive year TSMC has flagged, with an explicit four-year price ladder going through 2029. AI demand is the cited reason every time. And yet customer capacity bookings are still oversubscribed through the back half of the year. The 15% is being absorbed.

The reason is the CoWoS bottleneck nobody can route around. Advanced packaging capacity is the gating constraint on every frontier AI accelerator currently in production, and there is exactly one foundry running it at the volumes the AI labs need. The hyperscaler capex plans have committed to GPU and ASIC volumes that require those 3nm wafers and that CoWoS packaging through the end of 2027. There is no alternative supplier. Intel Foundry is recovering but not on the relevant scale yet. Samsung Foundry’s yield numbers on equivalent nodes are not where they need to be. So when TSMC quotes 15%, the customer signs the order.

The downstream pricing implication is the part nobody is saying out loud. If wafer cost is rising 15% on 3nm in H2 and another 5 to 10% the following year, and HBM cost is rising as the HBM4E ramp pulls forward, then the cost-per-token line on the back of every inference deck has to bend up at exactly the moment frontier labs are negotiating multi-year enterprise discounts. Either the labs eat the margin, or the labs renegotiate the discounts, or both. The “AI is getting cheaper every six months” graph that every keynote presenter uses for the next twelve months is going to need an asterisk.

tsmc3nmpricingnvidiaappleamdbroadcomcapacityai-hardware