Google has ordered more than three million of its custom TPU AI chips from Intel’s foundry, with production scheduled for 2028, according to people familiar with the agreement. The order is the largest external commitment Intel’s foundry business has landed since its turnaround began, and the news broke in public with a bang: Intel shares rose more than 11%.
The deal has been in negotiation for more than a year, the people said, and it survived several near-collapses over price, yield targets and scheduling. Google’s decision came down to a simple strategic need: its AI chip supply can no longer run exclusively through Taiwan Semiconductor Manufacturing Co., the world’s dominant chip foundry.
Google designs its own TPUs, tensor processing units, which power much of its AI training and inference, and the chips are among the highest-volume custom silicon in the industry. Until now, nearly all of them were made by TSMC. The Intel order gives Google a second source and a negotiating lever, and it gives Intel the reference account its foundry business has lacked.
For Intel, the deal is existential in scale. The foundry business, launched as part of a sweeping strategy shift in 2021, has burned cash and struggled to land large customers, and its most credible anchor was a promise rather than a contract. Google’s three million chips, made on Intel’s next-generation process nodes, convert that promise into something an investor can count.
The 2028 production date reflects the realities of chipmaking. Google needs time to design its TPU generations for a new process, Intel needs time to qualify the manufacturing line, and both companies need a buffer for the surprises that define leading-edge production. The people familiar with the deal said volume production is expected to ramp through 2028.
The chips will be made at Intel’s U.S. factories, the people said, a detail that matters in Washington. The U.S. government has spent billions to bring leading-edge chipmaking back onshore, and a deal that moves a major AI chip to American soil is exactly the outcome the policy was designed to produce.
For Google, the motivation is partly defensive. AI demand has TSMC’s advanced nodes booked out years in advance, and a single point of failure in a single country is a risk the company has quietly tried to diversify. The Intel order reduces that risk and, in negotiations with TSMC, gives Google a credible alternative to threaten.
TSMC’s response has been muted in public. The company said it remains Google’s largest chip supplier, which is true: the Intel order covers a fraction of Google’s total silicon, and most of its most advanced accelerators will stay with TSMC. But the psychological impact is real, and other foundry customers are watching how the first big defection plays out.
Analysts are split on the deal’s value. The financial terms weren’t disclosed, and the strategic value depends on yield and price, both unverified at this stage. The clearest read is the direction: Intel can now walk into any potential foundry customer’s office with a Google contract on the table, and in this industry, reference accounts are how orders compound.
The risks are the same ones that have haunted Intel’s foundry effort. Its process roadmap has slipped before, and a 2028 schedule gives room but no guarantees. If the production line misses its yield targets, the deal becomes a story about a missed opportunity rather than a turnaround, and the market that pushed the stock up 11% will push it back down faster.
For the broader chip industry, the deal cracks a structural assumption.
Intel has been preparing for this moment for years. It has been converting its U.S. factories for foundry work, a conversion that required new equipment, new process recipes and a new sales culture, and the company has said foundry revenue will grow sharply once its leading-edge nodes reach volume. The Google order is the first confirmation from a major customer that the conversion is real.
Google’s TPUs are the foundation of its AI operations. The company has shipped successive generations of the chips, each designed to run the models behind its search, cloud and Gemini products, and the Intel order covers chips from its coming generation, the people said. Google has long said it designs its own silicon to control cost and performance; the deal extends that control to manufacturing. For two decades, the most advanced AI chips were made by one company in one country, and everyone else’s strategy started with that fact. Google’s order says the assumption is negotiable, and other hyperscalers with custom silicon are now asking the same question Intel is answering.
For a company that spent a decade watching its manufacturing lead evaporate, the Google order is a vote of confidence denominated in millions of chips. Intel still has to make them, on time and at a profit, but for the first time in years, someone who matters is betting it will.
