The semiconductor supply chain may be entering its biggest realignment in years. According to reports, Google has tapped Intel's foundry to manufacture more than three million of its in-house tensor processing units in 2028, while Nvidia is evaluating Intel's most advanced 18A process for a next-generation graphics chip — two signs that the world's leading AI chip designers are quietly seeking an alternative to Taiwan's TSMC. Intel's stock jumped sharply on the news, but the story carries important caveats: nothing has been officially confirmed, and at least one major bank is skeptical of how much it really means.
For anyone tracking where AI hardware gets built, the useful move is to separate two things the headlines blur together: a real, durable trend toward diversifying away from TSMC, and a specific, still-unconfirmed transaction whose scope analysts dispute.
What Exactly Was Reported?
The reports originated with The Information on June 8, 2026, and were echoed by Reuters under the framing that Google and Nvidia are considering Intel as a "backup" chip manufacturer. Per the reporting, Google placed an order for more than three million TPUs to be produced through 2028 after months of validating Intel's advanced packaging. For scale, Morgan Stanley estimates Google will produce over six million TPUs across 2027 and 2028, which would make Intel's reported order roughly half of Google's 2028 output — a substantial win if it holds.
Separately, the reports say Nvidia is assessing whether Intel can manufacture a new processor that fuses four GPU dies into a single package, a design tied to Nvidia's next-generation Feynman architecture due around 2028. Nvidia has reportedly begun early testing on Intel's 18A process through multi-project wafer runs — a cost-sharing arrangement in which several customers validate small designs on one wafer — but has not committed to a production order. Investors reacted immediately, sending Intel shares up by high-single to low-double digits intraday and adding billions in market value, as Wall Street treated even unconfirmed AI-foundry wins as validation of CEO Lip-Bu Tan's manufacturing turnaround.
Why Is the Industry Moving Away From TSMC?
Whatever the precise scope of the Intel reports, they crystallize a genuine shift. TSMC manufactures virtually every leading-edge AI chip in the world, and it is straining to keep up — with the squeeze worst in advanced-packaging lines, the step that stitches multiple dies and high-bandwidth memory into a single module. TSMC's own leadership has warned the AI chip shortage will persist for years, and the company has been raising prices on its most advanced nodes.
That concentration has become a strategic risk boards can no longer ignore. Depending on a single company, in a single region exposed to geopolitical tension, for the chips at the center of a multitrillion-dollar AI build-out is exactly the kind of single point of failure that pushes large buyers to find a second source. Intel — long humbled in foundry — is the most credible Western alternative, and it has been building the case: its 18A node is already shipping in its own Xeon server processors, its EMIB advanced-packaging technology has reportedly reached yields near 90% with Google's TPU v8e (expected in late 2027) cited as a candidate to use it, and it has secured Tesla as the first major customer for its next-generation 14A process.
Why Are Analysts Skeptical?
Before declaring TSMC dethroned, the caveats deserve equal weight. JPMorgan poured cold water on the Google report, suggesting the TPUs would still be fabricated by TSMC with Intel handling only the packaging — a meaningful but far less transformative role than making the chips themselves. The bank reportedly called the excitement a "storm in a teacup."
Nvidia's engagement carries its own history. An earlier Reuters report indicated Nvidia had tested Intel's 18A process and "stopped moving forward," with the read that yield and performance fell short. The current evaluation has reportedly progressed beyond initial talks, but multi-project wafer runs are a long way from a committed order. Nvidia's current Rubin GPUs are built on TSMC's 3-nanometer process with CoWoS packaging, and earlier reports had its Feynman architecture slated for TSMC's A16 — so Intel would have to displace an incumbent already woven into Nvidia's roadmap. There is also an awkward fact: Intel itself remains a TSMC customer, with the compute die of its upcoming Nova Lake chip reportedly outsourced to TSMC's 2-nanometer process. The relationship is less a clean rivalry than a complex coopetition, and a few reported orders do not undo TSMC's structural lead in leading-edge volume.
What Makes Intel's Packaging the Real Prize?
The reason these reports center on packaging as much as on transistors is that modern AI chips are increasingly defined by how their pieces are assembled. As single-die designs hit the physical limit of how large a chip can be — the "reticle limit" — designers stitch multiple smaller dies into one package (Nvidia's reported four-die GPU is an example) and connect them with high-bandwidth interconnects. Intel's EMIB embeds a tiny silicon bridge between dies to carry signals at high density, and its reported yield gains there are part of what makes it attractive to designers who need advanced-packaging capacity TSMC cannot currently supply fast enough.
For Intel's 18A node, the foundry pitch rests on proving it can deliver leading-edge transistors at high yield and volume — the very thing that humbled Intel's manufacturing for the better part of a decade. That 18A is already shipping in Intel's own data-center chips is the strongest evidence the node is real; whether it can win and hold external customers at the scale of a three-million-unit order is the open question.
What Should You Watch Next?
Separate the trend from the transaction. The trend — large AI chip buyers diversifying away from total dependence on TSMC — is real, durable, and driven by genuine capacity and geopolitical pressure. The specific transaction — a confirmed, leading-edge three-million-TPU Intel manufacturing win — is reported but unconfirmed, and credible analysts believe Intel's role may be narrower than the headlines imply.
What would settle it: an official confirmation from Google or Intel specifying whether Intel is fabricating the chips or only packaging them; a committed Nvidia production order rather than evaluation wafers; and, beneath both, hard data on 18A yields at volume. Until then, the most accurate read is that Intel's foundry is gaining credibility at exactly the moment the industry most wants a TSMC alternative — and that the market is paying up for the possibility well ahead of the proof. This article is not investment advice.
Frequently Asked Questions
Did Google actually order 3 million TPUs from Intel?
It was reported by The Information and echoed by Reuters that Google placed an order for more than 3 million TPUs to be produced through 2028, but neither Google nor Intel has officially confirmed it. JPMorgan has cautioned that the chips may still be made by TSMC, with Intel handling only the packaging.
Is Nvidia switching to Intel from TSMC?
Not yet. Nvidia is reportedly evaluating Intel's 18A process and advanced packaging for a multi-die GPU tied to its Feynman architecture, running early multi-project wafer tests, but it has not committed to a production order. Its current Rubin GPUs are made on TSMC's 3-nanometer process.
Why are chipmakers looking for a TSMC alternative?
TSMC produces nearly all leading-edge AI chips and is straining to meet demand, especially in advanced packaging, while prices rise and a years-long shortage looms. Relying on one company in one region is a strategic risk, prompting buyers like Google and Nvidia to seek a backup such as Intel.
What is Intel's 18A process?
18A is Intel's most advanced manufacturing node and a centerpiece of its foundry comeback under CEO Lip-Bu Tan. It is already shipping in Intel's own Xeon server chips; the open question is whether Intel can win and reliably supply large external customers at high yield and volume.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
