TSMC Earnings on Deck: CoWoS, Micron's $250B Bet & China's Helium Ban

**Jamie Call** (0:00)
AI Hardware & Chips, Daily News.
I'm Jamie Call. Thanks for joining me.
Today, TSMC Earnings Rutel, the COOS update that moves markets.
TSMC's Earnings Call on July 16th is the single most important event in semiconductors this week, and the whole market knows it. The reason isn't revenue, it's COOS. That's TSMC's advanced packaging technology. And right now, it's the real bottleneck on AI accelerator supply, not logic fab output, packaging. CC. Wei's language on that call will either confirm the AI CapEx cycle is still accelerating, or introduce the first credible doubt. Either outcome re-rates the entire sector, Nvidia, AMD, Broadcom, the equipment makers. All of them move on whatever Wei says at 2 in the morning Eastern time.
Here's where things stand going in. COO's lead times currently stretch 52 to 78 weeks, which puts demand well into 2027 Previous guidance cited 2026 is the constraint window. The market expects that to slip. The key signal will be whether TSMC's customers are pushing out orders or doubling down. Those are very different stories, and Wei will know which one is true. Ahead of that call, TSMC has already made one significant move. The company informed major customers of price increases of 5 to 10 percent across its 7 manometer and below nodes. Nvidia, AMD and Broadcom can absorb that. Their margins allow it and they can pass costs downstream. The important distinction is what happens to mid-tier chip designers who don't have that buffer. For them, this is a margin squeeze with nowhere to go. The pricing move itself is a signal. TSMC doesn't raise prices without conviction about its own supply tightness. That's worth more than the number. The memory side of the AI infrastructure picture is moving on its own, and the scale is striking. Micron has committed $250 billion to expand HBM capacity onshore production in the United States and build out what it's calling friend-shored supply chains.
The strategic logic is direct.
Samsung and SK Hynix have held near total dominance in high bandwidth memory for years.
Micron is betting that the AI boom is permanent enough to justify breaking that duopoly through sheer capital commitment, with Chips Act incentives as the backstop.
There's the thing, the execution risk is real, and semiconductor history makes that clear. Overcapacity busts have followed aggressive expansion before. The key uncertainty isn't whether Micron can build, it's whether AI memory demand stays structurally elevated long enough to justify the spend, before a new memory architecture changes the equation entirely. The memory shortage isn't contained to data centers. PC shipments in the second quarter of 2026 fell 4.9% year over year to 62.8 million units. That's the first annual decline in two years, and the primary driver is memory cost pressure cascading from the HBM buildout into consumer DRAM and NAND. Relief isn't expected until at least the second quarter of 2027, and some analysts think that timeline is still optimistic. China's decision to ban helium exports is the newest variable in the supply chain picture, and it fits an established pattern. Beijing has now moved through rare earths, gallium, germanium and graphite. Halium is the latest addition. It matters because helium is a critical cooling material in semiconductor fabrication, and China controls roughly 90% of global helium imports despite holding modest reserves. The near-term disruption scope is still unclear. What's less unclear is the direction of travel.
Samsung's stock fell 7% on July 7th despite posting a solid earnings preview. The institutional selling is the more telling fact. After gains of over 200% in the first half of 2026, investors appear to be questioning whether memory prices can hold through the second half. That's a different concern in current fundamentals and it's worth watching. Consider this. On the demand side, Microsoft's Fairwater facility in Mount Pleasant, Wisconsin went live on June 23rd. It's a purpose-built AI data center adding meaningful GPU capacity to Azure's US North Central region. A second phase, called Lakeview, is already under construction for 2027 Alongside similar commitments from Amazon, Alphabet and Meta, the hyperscader CapEx picture still points to sustained AI infrastructure demand. Roughly $700 billion in collective cloud spending in 2026, most of it flowing through TSMC. The near-term watchpoint is straightforward. TSMC's July 16th call will either validate or complicate everything else covered today.
If Wei signals CoWOs as sold out well into 2027 with no customer pushouts, the AI CapEx cycle stays intact and the sector re-rates upward. If the language softens, the market will look for confirmation that demand is peaking. Samsung's trajectory and Micron's execution will be the secondary signals to track through the rest of the quarter. Thanks for listening. This podcast was built using AI technology, a YesWe production.