Google orders chips from Intel and Nvidia is testing its tech, as TSMC’s grip on AI starts to strain

For years, the AI boom has run through a handful of factory floors in Taiwan. That dependence is starting to look risky enough that even Nvidia and Google are shopping for a backup, and the unlikely name on the list is Intel.

Google has placed an order with Intel to manufacture more than three million of its in-house tensor processing units in 2028, while Nvidia is evaluating Intel’s advanced packaging and its most cutting-edge 18A process for future chips, The Information reported on Monday, citing four people with direct knowledge of the talks. Intel’s shares jumped about 12 per cent on the news.

The driver is scarcity.

TSMC, the Taiwanese giant that manufactures virtually every leading-edge AI chip, is straining to keep up with demand, with the squeeze worst in the advanced-packaging lines that stitch chips and memory together. For the companies designing the world’s most sought-after silicon, relying on a single supplier in a single country has become a strategic liability.

The two approaches differ in seriousness.

Google’s is a firm order: more than three million TPUs in 2028, after months of testing Intel’s packaging, part of a build-out Morgan Stanley estimates at over six million TPUs across 2027 and 2028.

Nvidia has not committed. It is running early trials, including multiproject wafer runs on 18A, and testing whether Intel can build a processor that fuses four graphics chips into one, a design tied to its Feynman GPU architecture due in 2028.

For Intel, even cautious interest is a milestone.

The one-time leader of American chipmaking has spent years trying to turn its contract-manufacturing arm into a credible rival to TSMC, with limited success and heavy losses. It has been courting Apple as a foundry customer, and both the US government and Nvidia have taken equity stakes in the company.

Actual orders, rather than goodwill, are what the turnaround has lacked.

The near-term opening is in packaging rather than the most advanced chipmaking, where Intel still trails. SK Hynix is reportedly testing whether its high-bandwidth memory works reliably with Intel’s packaging too. Whether 18A can match TSMC on yield, the share of chips that come out usable, remains the open question that has tripped Intel up before.

It also underscores how far Google has gone to control its own silicon. Its tensor processing units now ship in the millions, and it has been spreading orders across a roster of partners to reduce its reliance on both Nvidia and TSMC. Adding Intel is one more hedge.

None of this dethrones TSMC, whose lead in leading-edge manufacturing remains vast. But it is a sign that the AI industry’s most powerful buyers no longer want all their eggs in one Taiwanese basket, and that Intel, written off by many, may yet have a role to play in supplying them.

For a company that has needed a customer to believe in its comeback, Google just became one.