CoWoS at 260K, EMIB-T & Broadcom's $230B AI Forecast

Jamie Cole (0:00)

AI Hardware & Chips, Daily News.

I'm Jamie Cole, thanks for joining me.

Today, the packaging crisis is real now.

Advanced packaging is now the binding constraint on AI chip shipments, not transistor density, not logic node maturity, packaging. TSMC is projecting it will need to double its COWAS capacity from 130,000 to 260,000 monthly equivalent wafers by the end of 2028, and the industry is scrambling to keep pace. That number matters, because COWAS, TSMC's chip-on-wafer-on-substrate technology, is what allows AI accelerators to integrate high-bandwidth memory at the densities Nvidia and others actually need. You can have the most advanced logic node in the world. If you can't package it efficiently, you can't ship it at scale. That's the situation the industry is in right now. TSMC has officially confirmed it will move to 14 times Reticor COO's production in 2028 That's a package integrating roughly 10 compute dies and 20 HBM stacks in a single unit. The signal here is that TSMC isn't just adding volume, it's expanding package density. Both dimensions matter, and the industry hasn't fully priced in what that combination means for AI accelerator design.

The 260,000 wafer figure is an analyst projection, not official TSMC guidance. That distinction is worth holding on to. Yield issues or demand shifts could push the ramp timeline out.

But the reticle confirmation is official, and it anchors the direction clearly. The packaging crunch is opening a door for Intel. Intel's EMIB-T technology, a competing approach to high-Bamblet-Chipler integration, is projected to reach 40,000 to 45,000 monthly equivalent wafers by 2028 Google has already reserved capacity there for more than 3 million TPUs. The important distinction is that Intel isn't winning on logic node performance. It's winning on packaging availability. That's a meaningful repositioning.

Intel Foundry's credibility on logic has been questioned repeatedly, but EMIB gives it a viable entry point precisely because the market is desperate for alternatives to TSMC. Here's the thing. ASE and Amkor are also being qualified by hyperscalers. Combined, they're projected to reach 110,000 monthly equivalent wafers by 2028 Cloud providers are actively building a multi-supplier packaging strategy. The TSMC's single-vendor dependency is being hedged deliberately and at scale. On the logic side, the iPhone 18 Pro launch with the A20 Pro chip confirmed TSMC's 2nm process is commercially viable at volume. The A20 Pro runs dual 16-core neural engines, doubling AI processing versus the A19 Pro. Previous uncertainty around 2nm yield and ramp timing is now resolved. The other thing the A20 Pro signals is strategic. TSMC is integrating system-level design decisions, not just fabricating nodes. Off-thermal path memory placement in the A20 Pro package shows TSMC thinking beyond the wafer. That has implications for how future data center chips get designed, not just smartphones. TSMC's revenue concentration is also shifting in a way that tells you something structural. Apple's share of TSMC revenue has dropped from 25% in 2023 to 19% in 2025 The second largest customer, Nvidia, has risen from 11% to 17% in the same window. Consider this. Two years ago, Apple and Nvidia were 14 percentage points apart. Now they're two. That compression reflects where AI accelerator demand is pulling TSMC's most advanced capacity. Apple still matters, but Nvidia's trajectory is the one reshaping TSMC's customer calculus.

Broadcom's numbers deserve attention and some scrutiny. The company that guided AI semiconductor revenue to 58 billion in fiscal 2026, 115 billion in 2027, and 230 billion in 2028 Custom XPU shipments made up 73% of its AI revenue in the most recent quarter.

The key implication is that Broadcom is no longer a networking and connectivity company, with some AI exposure.

Its positioning is a full AI infrastructure player. Whether 230 billion in 2028 is realistic depends heavily on hyperscaler CapEx continuing to grow. If model efficiency improves faster than expected, or if cloud spending cools, that trajectory congresses quickly. Treat the forecast as a directional signal, not a guaranteed outcome. The two metrics worth tracking over the next 12 months are COWO's capacity delivery against the 2028 target, and whether Intel's EMIB-T qualifications convert into real volume at scale. Intel's Foundry track record on hitting capacity commitments is uneven. The Google TPU deal is real. Execution at 40,000 wafers monthly is still an open question. The packaging layer is now where the AI hardware race is actually being decided. That's the frame to carry into every chip story you read from here. Thanks for listening. This podcast was built using AI technology. A Yes We production.