Intel's Packaging Moat Crumbles, GDDR Shortage Bites & CoWoS Booked to 2027

**Jamie Cole** (0:00)
Oh, Hardware & Chips, Daily News.
I'm Jamie Cole. Thanks for joining me. Today, TSMC copies Intel's playbook while closing FoundryGap. I am an actor, Jamie Sinlu, and good friends of family.
TSMC is developing a Silicon Bridge packaging technology that directly mirrors Intel's EMIB, and Nvidia's already evaluating it for future processes. That's the clearest signal yet that Intel's last remaining differentiation of the Foundry race is being closed, fast. EMIB, Intel's embedded multi-die interconnect bridge, has been one of the few genuine technical talking points Intel Foundry could offer chip designers that TSMC couldn't match. That advantage may now have a shelf life measured in quarters, not years. TSMC is building the capability with packaging partner Kinsus, and the timing matters. Intel's Foundry business is still deep in the red, posting a $2.1 billion operating loss last quarter, with external revenue of just $293 million, roughly 5% of total Foundry sales, despite 31% growth. The technical moat was never going to carry the business alone, but losing it removes a concrete reason for designers to look seriously at Intel's process nodes. Intel did move aggressively on a different front. The company repurchased its 49% stake in Fab 34 from Apollo for $14.2 billion, reclaiming full control of its leading European facility in Ireland. The signal there is commitment over caution. Intel is doubling down on capacity at the exact moment its 18A process enters risk production and its external customer pipeline remains thin. Apple is still in discussions, not committed. The important distinction is whether this is strategic conviction or sunk cost momentum. The answer probably depends on whether 18A yields hold. The memory supply picture is getting worse and it's spreading. Nvidia moved first in late July, raising kit prices 10-15%.
AMD confirmed its following, with Radeon price increases in August. Gddr6 and 7 memory costs have tripled since late 2025, and the supply ceiling is set by three DRAM makers. Neither AMD nor Nvidia controls that lever. The signal here isn't competitive pricing strategy.
It's a coordinated response to a structural shortage neither company can solve unilaterally. The contagion is already broader than GPU cards. Console makers are raising prices citing the same supply pressure. Relief looks unlikely before 2028
TSMC's advanced co-OS packaging is fully booked through 2027 The 14 reticle package roadmap, targeting 2028, would integrate around 10 compute dies and 20 HBM stacks per package. That's a substantial density leap. What matters in practice is that TSMC is consolidating both the capacity bottleneck and the capability roadmap simultaneously. The EMIB-like development with Kinsus fits the same logic. TSMC isn't just filling orders. It's building a packaging moat that closes off the alternatives.
The deployment numbers are striking. The world is on pace for roughly 200 million H100 equivalent ships by end 2028, doubling every nine months from a current base of around 20 million. Data centers in Iowa, Wisconsin and elsewhere will be more than 100 times more powerful than the hardware that trained the original ChatGPT.
Global data center power demand sat at 64 gigawatts in 2025 and is expected to exceed 256 gigawatts by 2030 That exceeds the combined electricity consumption of South America and Africa. US companies, led by Amazon, Google, Microsoft, Meta and Oracle, are projected to spend around $750 billion this year on AI infrastructure alone. Here's the thing, the unresolved questions are the ones worth tracking. Whether Intel secures a marquee external foundry customer before 18A moves to volume production. Whether TSMC's EMIB-like capability actually lands with Nvidia on a committed design. And whether Gddr6 supply relief arrives anywhere close to current timelines, given that three companies control 90-95% of production and the most optimistic forecasts push relief past 2028
The packaging race is now the core competitive story in AI chips. Whoever controls advanced interconnect at scale controls the bottleneck. That's the frame to hold on to. Thanks for listening. This podcast was built using AI technology, a yes we production.