TSMC Earnings, Meta Iris & the CoWoS Crunch That Defines 2026

**Jamie Cole** (0:00)
AI Hardware & Chips, Daily News. I'm Jamie Cole. Thanks for joining me. Today, the Co-O's Crunch that never ends. TSMC just announced two new advanced packaging plants in Qiai, and the Co-O's Crunch is still not going away. That's the clearest signal from this week's data. Even with those new facilities, the supply-demand gap for Co-O's capacity is only expected to narrow from roughly 20% to 10% by year-end. Packaging, not chip fabrication, is now the binding constraint on AI accelerator shipments.
That's worth sitting with for a moment. The background here matters. Co-O's demand is projected to hit 1 million wafers in 2026 up from 370,000 in 2024
TSMC is expanding capacity at around 80% annually. It still won't be enough. The first Qiai facility entered mass production in June, which is real progress.
But the structural gap persists and it's now confirmed to run through at least 2026
The capacity allocation story inside that bottleneck is where the competitive dynamics get shot. Nvidia holds roughly 60% of total co-os output, around 595,000 wafers, and has already secured more than 50% of the 2026 and 2027 expansion. AMD sits at around 11%, about 105,000 wafers, and it's scaling both CPUs and GPUs through that same constrained pipe simultaneously. That's a structural disadvantage AMD can't easily trade its way out of, and it's trading near a 70 times forward earnings multiple. A minor packaging miss would reset that valuation quickly. TSMC posted an all-time revenue record in quarter two, approximately $1.27 trillion new Taiwan dollars, with N3 and CoWAS both sold out through year-end. The signal here isn't just the number, it's that TSMC is capacity-constrained, despite record expansion spending. Thursday's earnings call on July 16th is the next real proof point. If management signals a HAPEX cut, that tells you they're seeing demand soften. If they hold or raise, the AI infrastructure cycle is intact. Watch the CAPEX guidance, not the headline revenue. The more structural pressure on Nvidia is coming from its largest customers. Meta is accelerating its Iris custom chip to production in September, targeting a doubling of compute capacity and reducing dependence on third-party GPUs. Alphabet is now commercializing its TPU chips through a NeoCloud offering, selling capacity externally rather than keeping it proprietary. These moves are deliberate. Hyperscalers are just hedging against Nvidia. They're building parallel supply chains. The displacement rate of Nvidia GPUs and each hyperscalers workload remains unclear, but the direction is not. Further out in the supply chain, Intel is committing 5 billion euros to expand its Leixlip campus in Ireland, focused on Xeon data center processes and Foundry customer support.
Samsung is running an aggressive hiring push across HBM Core die, Base die, reliability and packaging engineering to scale HBM4 and 5 production. Both moves reflect a broader bet that Nvidia's supply chain vulnerability is real and durable enough to justify heavy capital commitment now.
Here's the thing. One risk that's harder to hedge landed Monday. US Iran tensions pushed oil prices higher and semiconductor and AI stocks sold off sharply. Nvidia retreated to the $200 support level. Energy cost is no longer a background assumption for data center economics. It's now actively priced as a material risk to infrastructure ROI through 2026 and beyond. The near term watchpoints are clear. TSMC's earnings Thursday, Meta Iris production confirmation in September, and whether energy prices stabilize or keep climbing. Those three will tell you more about the next phase of this cycle than almost anything else.
Thanks for listening. This podcast was built using AI technology, a YesWe production.