0

Alphabet moves to diversify supply chain with Intel amid TSMC capacity crunch

In sum – what we know:

• A reported foundry win – Google has reportedly placed a firm order for over 3 million custom TPUs through Intel Foundry for 2028, though no company has confirmed it.
• Diversification, not defection – Intel would handle roughly half of Google’s projected 6-million-unit volume across 2027–2028, with TSMC keeping the rest.
• A bet on Intel’s execution – The order hinges on Intel hitting aggressive yield and timeline milestones on its 18A node, where it has a history of delays.

Alphabet’s Google has placed a firm order with Intel to manufacture more than 3 million of its custom Tensor Processing Units (TPUs) for 2028, according to a report from The Information citing people with direct knowledge of the discussions. The chips would be produced through Intel Foundry, the contract manufacturing arm Intel has spent years building in hopes of challenging TSMC.

It’s worth being clear about what we know and what we don’t. Neither Google, Intel, Nvidia, nor TSMC has officially confirmed the deal, and the specifics — volume, process node, fab locations — all trace back to anonymous sources. 

Still, if the reported numbers hold up, this would be one of the largest single external foundry wins in Intel’s history. And the timing isn’t a coincidence — it lands squarely in the middle of a capacity crunch at TSMC that has the entire AI industry scrambling for alternatives.

Google’s TPU strategy and production split

Google is projected to deploy more than 6 million TPUs across 2027 and 2028, a figure that says a lot about the scale of its AI ambitions. Those chips power internal workloads like search and YouTube recommendations, the Gemini family of models, and the Cloud TPU instances Google sells to enterprise customers as an alternative to Nvidia GPUs. The TPU has long been central to how Google differentiates its cloud business — TPUs are custom silicon it controls end to end, rather than off-the-shelf parts everyone else can buy.

Intel is expected to handle roughly half of that 6-million-unit volume, with TSMC likely retaining the rest. That split matters. This is diversification, not a defection — TSMC has fabricated every generation of Google’s TPUs to date, and there’s no indication that relationship is ending. 

The logic is straightforward. A multi-foundry strategy reduces Google’s dependence on a single manufacturing partner, improves resilience against regional disruptions, and — perhaps just as importantly — gives Google negotiating leverage on pricing and priority at both foundries. Once TSMC knows its biggest customers have a qualified second source, the dynamics of every future capacity conversation change.

Intel’s technological appeal and validation

This order didn’t come out of nowhere. Google reportedly spent months qualifying Intel’s advanced packaging technologies — the likes of EMIB for die-to-die interconnect and Foveros-style 3D stacking — before committing volume. For modern AI accelerators, packaging is arguably as important as the process node itself, because these chips combine large compute dies with stacks of high-bandwidth memory (HBM) in increasingly complex configurations. Get the packaging wrong and the world’s best transistors don’t matter much.

The 2028 timeline also lines up neatly with Intel’s roadmap for its 18A process node, which the company has been pitching as competitive with TSMC’s leading edge by the late 2020s. To be clear, Intel hasn’t confirmed which node will produce the TPUs, and the reporting doesn’t conclusively say whether Intel will handle both wafer fabrication and packaging for all 3 million units. 

For Intel, this is the validation its “IDM 2.0” foundry strategy has badly needed. The company has poured billions into new fabs and process development while struggling to land marquee external customers — and a firm, multi-million-unit commitment from one of the world’s largest cloud providers is exactly the proof point it can take to the likes of Nvidia, Apple, and Amazon. Nvidia, notably, is reportedly already testing 18A and Intel’s packaging for future GPU architectures, though it hasn’t placed a firm order. Investors got the message regardless; Intel’s stock jumped roughly 12-13% on the news.

That said, a 2028 order is a bet on execution, not a guarantee of it. Intel has to hit aggressive timeline, yield, and reliability milestones on a leading-edge node — territory where TSMC has years of proven track record and Intel has a history of painful delays. If Intel slips, Google goes back to fighting for TSMC capacity along with everyone else.

TSMC capacity constraints and geopolitical factors

None of this should be read as TSMC losing its grip on the high end. The Taiwanese foundry remains the dominant manufacturer of leading-edge AI chips, producing the overwhelming majority of Nvidia’s GPUs and most of the custom accelerators designed by hyperscalers. The problem isn’t demand — it’s supply. TSMC’s leading-edge capacity is effectively sold out through at least 2027-2028, and demand from the AI boom has vastly outstripped its ability to add wafer starts and advanced packaging. Even its upcoming Arizona fab is reportedly booked before it’s finished.

So Google’s move reads less like a vote against TSMC and more like a natural consequence of demand outpacing what any single supplier can build. The risk for TSMC is longer-term. Once customers go through the expensive, months-long process of qualifying an alternative foundry, more business tends to follow — and the near-monopoly TSMC enjoys at the very high end starts to erode at the margins.

There’s a geopolitical dimension here, too. Concentrating the world’s advanced chip manufacturing in Taiwan carries real risk given ongoing China-Taiwan tensions, and that’s not a comfortable position for a company planning its AI infrastructure years in advance. Intel’s expanding fab footprint in the U.S. and Europe, backed by the CHIPS Act and various European subsidy programs, gives Google a way to friend-shore a meaningful chunk of its silicon supply chain. For a politically scrutinized company, manufacturing critical AI hardware on American soil isn’t a bad story to be able to tell, either.

Whether any of this materializes as reported won’t be clear for a while.