TSMC's Packaging Gambit, Google TPU9 & AWS's $220B HBM Bill

**Jamie Cole** (0:00)
AI Hardware & Chips, Daily News. I'm Jamie Cole, thanks for joining me.
Today, packaging wars.
TSMC stock hit its daily limit up in Thailand this week after the company confirmed it's developing its own alternative to Intel's EMIB packaging technology. That's the signal worth tracking. Because what it tells us is that TSMC no longer has the luxury of ignoring Intel in advanced packaging. Coos, the technology that's been the default home for every major AI accelerator, is running out of room. The capacity problem is straightforward. Coos is fully booked through 2027 That's not a near-term blip, it's a multi-year constraint, and hyperscalers designing next-generation chips can't afford to wait. Intel's EMIB solves the same core problem differently, embedding silicon bridges directly into the substrate rather than using a large interposer. It's lower cost, smaller footprint, and increasingly, it's available. The customer that makes this a real crisis for TSMC is Google. TPU Generations 7 and 8 both ran on CoWAS-L. The 9th generation is reportedly being evaluated on Intel's EMIB instead. That's not confirmed, but the direction is credible, and the implication is significant. If Google locks a TPU design into Intel's packaging ecosystem, that's not a one-quarter revenue swing. Custom chip designs are sticky, switching costs are high. Once a layout is optimized for a specific packaging architecture, you don't revisit it lightly. That's the strategic weight behind TSMC's move. The quasi-IEMID partnership with substrate maker Kinsus isn't just a technical hedge, it's a retention play. TSMC is telling its customers, you don't have to go to Intel to get a bridge-based alternative. The question is timing. Intel has been shipping EMIB commercially for years. TSMC hasn't disclosed a production schedule for its version, and Kinsus will need to prove it can scale yields before any hyperscaler commits a design to it. Intel's position here is genuinely stronger than it looked 18 months ago. CEO Pat Gelsinger has described EMIBT demand as very high, with high-volume production scaling through 2027
Unimicron is bringing EMIBT substrate mass production online in 2027 at around 50% initial yields. That yield figure is the one to watch. 50% is workable at launch, but it needs to climb fast for Intel to meet the backlog it's signalling. The Foundry business overall is still deeply in the red. Intel Foundry posted a $2.1 billion operating loss in the second quarter of 2026
That's painful, but the trajectory matters.
It improved $348 million sequentially from the first quarter, driven by yield and cycle time gains. The loss is narrowing. Whether it narrows fast enough to be sustainable is still unresolved. Here's the thing. Away from the packaging battle, MediaTek is making a move that deserves attention. Its board approved a $5 billion discretionary budget targeting the data center custom chip market.
The specific asset here is a 448GB SerDes interface that's production ready in the second half of 2027
MediaTek is pointing directly at Google's TPU supply chain and Broadcam's custom ASIC dominance. The target is 15-20% of what's currently a roughly $2.5 billion market by 2028 Ambitious is understating it. Broadcam is deeply embedded, but the funding is committed and the SerDes timeline is firm. The infrastructure spend underpinning all of this keeps escalating. Amazon raised its CapEx guidance to $220 billion, up from $200 billion, explicitly citing HBM cost inflation as the driver.
The $20 billion increase is the price of a memory market that can't keep pace with AI demand. AWS remaining performance obligations hit $496 billion, up $132 billion in a single quarter.
That's multi-year demand locked in. The supply chain has to deliver against it.
The risk that's building quietly underneath all of this is substrates. Both TSMC's quasi-EMIB and Intel's EMIBT depend on advanced IC substrates. Kinsus and Unimicron are racing to supply two competing packaging futures simultaneously. If substrate capacity becomes the binding constraint in 2027, it doesn't matter which packaging technology wins the customer. Neither scales. The near-term signals to watch are precise. Does Google formally commit TPU9 to Intel EMIB? Does TSMC disclose a production timeline for its quasi-EMIB alternative? And can Unimicron push EMIB-T yields meaningfully above 50% before demand overwhelms its capacity? Those three data points will tell you more about where the packaging more settles than any headline number. Thanks for listening. This podcast was built using AI technology. A YesWe production.