The one packaging monopoly cracks

The chips that train AI models are not single slabs of silicon. A finished accelerator is a compute die lashed to stacks of high-bandwidth memory, and the lashing-together — called advanced packaging — is its own hard manufacturing step. For roughly a decade one company, TSMC, has effectively owned the only version of it that works at scale, a process called CoWoS. That step, not the chips themselves, is now the bottleneck on how many AI accelerators the world can build: CoWoS is booked solid through 2027, with NVIDIA alone holding more than half of 2026-27 capacity.

A monopoly that shaped the entire AI hardware supply chain now has a second door.

So Google went looking for a second source. It reportedly booked Intel's rival packaging process, EMIB, for a large share of its 2028 custom AI chips — over three million of them by one account, though neither company has confirmed it. What makes Intel credible for the first time is the economics: EMIB embeds tiny silicon bridges only where two dies actually touch, rather than seating everything on one wasteful slab of silicon, which analysts at Bernstein estimate costs a few hundred dollars per chip against $900-1,000 for CoWoS on a top-end part.

The headline oversold it. Intel's stock jumped about 11% on a 'three million chip order,' and then JPMorgan pointed out that the chips themselves are still fabricated at TSMC — Intel only does the back-end packaging. The market had priced a foundry comeback for what is a packaging contract. But the structural fact survives the correction, and it is the one that matters: the single supplier everyone leaned on can no longer make enough, and even its biggest customer is hedging. NVIDIA is quietly test-running Intel's process to fuse four dies for its next GPU, and SK hynix is testing it for memory. A monopoly that shaped the entire AI hardware supply chain now has a second door.