Google orders 3M+ AI chips/TPUs from Intel for 2028

Strip away the stock pop and the story underneath is structural. For years, essentially every leading-edge AI accelerator on Earth — Nvidia's GPUs, Google's TPUs, Amazon's and Meta's custom ASICs — has been etched in TSMC's fabs. That concentration is now the binding constraint: TSMC is straining to keep up with AI demand ^[1], and when a single supplier can't take more orders, the only lever a hyperscaler has left is a second source. Google reportedly pulling the trigger on more than 3 million TPUs from Intel for 2028 is significant less for the unit count than for who placed it ^[1]. A hyperscaler designing its own silicon is the most demanding, most sophisticated foundry customer there is, and Intel landing one — after Google spent months testing its manufacturing technology ^[2]— is the first real external validation that Intel Foundry can be trusted with bleeding-edge AI parts. Morgan Stanley's estimate that Google's total TPU output will top 6 million units across 2027 and 2028 frames how large the pie is ^[2]; even a minority slice routed to Intel is a meaningful diversion from TSMC's order book.

Here is where the bullish read needs a caveat. There is a world of difference between Intel assembling a chip and Intel manufacturing the transistors inside it. Several close observers — including a well-followed semiconductor leaker and commenters in Intel investor circles — have pushed back on the framing, arguing the 3-million-unit TPU order may be for Intel's advanced packaging (fusing and connecting silicon, e.g. EMIB-class interconnect) rather than fabbing the dies themselves on Intel's process node. That matters because packaging, while lucrative and capacity-relevant, does not prove Intel's 18A node is competitive with TSMC's most advanced lines. The genuinely node-level signal sits on the Nvidia side: Nvidia is running preliminary multiproject wafer validation on Intel's 18A process, testing whether Intel can fuse four graphics chips into one package for its 2028 Feynman architecture ^[1][2]. Crucially, Nvidia has not ordered anything yet. So the cleanest way to read today's news is two separate signals braided together — a Google volume commitment whose exact scope (package vs. wafer) is unconfirmed, and a Nvidia process evaluation that is real but still pre-commercial.

The contrarian case is worth stating plainly. None of this is officially confirmed — Google, Nvidia, and Intel all declined to comment, and the underlying claim rests on The Information's sourcing ^[2][3]. Skeptics in Intel investor forums went further, arguing parts of the narrative recycle earlier 2026 reporting on Intel-Google collaboration, and that the only genuinely new elements are the 18A node confirmation and the concrete 3-million-unit 2028 volume. Yet the timing is exactly what Intel's turnaround thesis required. Its foundry arm posted $5.4 billion in Q1 2026 revenue that is still overwhelmingly internal demand ^[3]; the missing ingredient has always been a marquee external customer. Analysts captured both the promise and the peril: TECHnalysis Research's Bob O'Donnell says the real turnaround signal is foundry becoming significant by 2027 ^[3], while Cerity Partners' Michael Schulman calls the whole effort a high-stakes gamble to become a credible rival by 2030 ^[3]. A roughly 10-12% one-day surge to a record high ^[1][3]prices in the optimistic ending of a story whose first chapter is still a set of anonymous sources.