Google is giving Intel something it has badly needed: a serious external customer for advanced AI chip manufacturing.
Intel has spent years trying to convince the market that its foundry business can be more than a costly turnaround story. Google may have just made that argument easier to believe. According to The Information, Google has placed an order for Intel to manufacture more than three million Tensor Processing Units for 2028 production, after testing Intel’s advanced packaging technology.
That is not a small trial balloon. TPUs are the custom AI chips Google uses to train and run models across its own products and Google Cloud, and the company has been pushing them harder as customers look for alternatives to Nvidia GPUs. The order reportedly comes as Google is expected to produce more than six million TPUs across 2027 and 2028, based on Morgan Stanley estimates cited in the report.
For Intel, the timing matters as much as the volume. The company has struggled to turn its manufacturing arm into a credible rival to TSMC, while its own AI accelerator products have failed to command the same attention as Nvidia’s GPUs or Google’s TPUs. A hyperscaler order of this size gives Intel a more tangible role in the AI infrastructure boom, even if it does not yet prove that the foundry strategy has fully turned.
The foundry business has been Intel’s great promise and its great burden. Pat Gelsinger relaunched the external manufacturing push in 2021, but Intel then ran into delays, heavy losses and investor impatience. Lip-Bu Tan inherited a company that still has valuable engineering depth, but also one that must show customers can trust it with high-volume production.
Google’s reported order helps because it is tied to a real bottleneck. TSMC remains the dominant manufacturer of advanced chips, but its leading-edge wafer capacity and advanced packaging lines have been under intense pressure from AI demand. Nvidia, Apple, AMD, Broadcom, Google and others all need access to the same limited manufacturing ecosystem. When supply gets tight enough, even cautious chip designers start looking for second sources.
Intel’s opening appears to be packaging first. The company’s EMIB technology connects multiple chip components inside a package, taking a different approach from TSMC’s CoWoS system. That matters because modern AI processors are no longer just about the main compute die. They rely on high-bandwidth memory, fast interconnects and tightly integrated chiplets. If packaging is constrained, the whole roadmap slows down.
Still, this is not a victory lap. Customers often test alternative suppliers without shifting their most sensitive designs at scale. Intel must prove yield, reliability and delivery discipline, not just win headlines. Nvidia is also reportedly evaluating Intel technology for future multi-chip GPU designs, including work tied to its 2028 Feynman architecture, but it has not placed a manufacturing order. That distinction matters.
Google is building leverage into its supply chain
Google’s decision should also be read as part of a larger strategy. The company is not simply buying chips. It is building leverage across the entire AI compute stack. It has its own TPUs, a growing Google Cloud business around those chips, design relationships with Broadcom and other silicon partners, and reported outside compute arrangements where it needs additional capacity.
That is the real lesson for founders and investors watching hyperscaler spending. The AI infrastructure market is moving away from a single-vendor story. Nvidia remains central because its GPUs, networking and CUDA ecosystem are still the default for many training workloads. But the biggest cloud companies are rich enough, technical enough and motivated enough to build custom silicon where the economics make sense.
Inference is where this becomes especially important. Training a frontier model is expensive, but running that model for millions or billions of users can become the larger recurring cost. Google has spent more than a decade refining TPUs for exactly that kind of high-volume workload. If it can produce them in greater numbers and place them behind Google Cloud services, it gains a pricing and supply advantage that smaller AI companies cannot easily copy.
That does not mean Nvidia is suddenly vulnerable in a simple way. Developers still value software maturity, broad compatibility and the ability to move quickly. But the buyer side of the market is changing. Hyperscalers are no longer just competing for Nvidia allocation. They are trying to shape the supply chain itself, from chip design to packaging to cloud rental models.
For Intel, the next test is whether Google becomes the start of a customer pattern or remains an exception created by TSMC scarcity. A single large order can lift the stock and improve sentiment, but foundry credibility is built through repeated execution. If Intel can deliver millions of advanced AI chips on time in 2028, the company will have a stronger case that it belongs in the next phase of AI infrastructure. Until then, the market will watch the order book, the yields and whether more hyperscalers decide that Intel is no longer just a backup plan.
Also read: A Security turns AI hacking into a $37 million venture bet • Bending Spoons is testing Wall Street with a planned US IPO • SoftBank's AI selloff shows investors are testing the boom
