Microsoft is the first company to say publicly that it will use Intel's EMIB advanced packaging inside its own first-party Azure chips. Beyond that, the joint announcement of 18 November is remarkably quiet.
No Azure part was named. No manufacturing node, no packaging order size, no volume production date, no contract value, no purchase quantity. What the two companies did confirm was a packaging technology and a multi-year commitment to Intel Xeon server CPUs that extends to the next generation, Diamond Rapids.
That asymmetry deserves more attention than the headline.
The bridge, not the interposer
EMIB solves a physical problem differently from the 2.5D packaging most AI accelerators use. Instead of placing a large silicon interposer beneath the dies, Intel embeds a local high-density silicon bridge in the package substrate and routes wiring only where adjacent dies meet. The savings are in the silicon dedicated to interconnect.
A full interposer has to span every die in the package and carry wiring for all of them. A bridge is placed only where two dies need to talk, so cost scales with the number of boundaries rather than with package area.
Intel says the approach has been in volume production in its own products since 2017, across server CPUs, GPUs and Altera FPGAs. That history matters, because the obvious objection to any new packaging supplier is whether the process has already shipped at scale.
Two variants extend it. EMIB-M adds MIM capacitors for voltage stability. EMIB-T adds through-silicon vias for vertical power delivery, aimed at chips placed beside HBM.
On 29 July 2026, Intel said its packaging could reach 8 times the industry standard reticle limit. The claim points at the same constraint from the other side: once a package can hold that much silicon, the limit on how large a part grows stops being the lithography field and becomes the wiring between the pieces. The American hub is Fab 9 at Rio Rancho, New Mexico, with 2,700 employees and 500 suppliers.
What Microsoft is and is not moving
The Maia lineage sets the context. Maia 100 and Maia 200 were built at TSMC with CoWoS-S packaging. Maia 300, announced in October 2026, is a multi-die design whose compute die uses TSMC N2P.
So the EMIB decision does not move Microsoft's leading-edge logic. The compute die stays where it was. What changes is that assembly and die-to-die interconnect gain a second lane, at a layer Microsoft does not control and cannot easily duplicate.
The architecture change makes that lane more valuable than it would have been two years ago. A multi-die part turns packaging from a fixed cost into a decision taken once per product, and it makes the die-to-die link a performance variable rather than a footnote.
The Xeon half
The processor side of the deal is more concrete, and it is where the measurable numbers sit. Microsoft confirmed multi-year Xeon supply capability covering Diamond Rapids.
The co-engineering claims are narrower but specific. CXL combined with Xeon 6 Flat Memory Mode gives Azure M-series v4 virtual machines up to 2.3 times the performance on certain workloads. Azure is the first cloud to enable Intel QuickAssist in virtual machines, with SQL Server backups up to 2.5 times faster. TDX protects the memory and register state of confidential virtual machines when they migrate across hosts.
Those figures are workload-specific and supplied by the companies. They describe tuning, not a platform shift.
Packing the two commitments into one release is itself informative. The packaging relationship arrives attached to a processor commitment with a multi-year horizon, which gives Intel a reference account in the part of its business that is not in question.
The 18A question stays open
On the same day, Intel chief executive Lip-Bu Tan said the company had signed its first external customer for the 18A process, without naming it. CNBC reported the remark. An Intel technology executive said the customer signed the 18A-P variant, which entered risk production in June.
Nothing in the Microsoft announcement connects the two. The temptation to join them is exactly the inference the disclosure does not support.
Why the silence is the signal
A packaging order can be small and still be strategically large. The withheld items are precisely the ones that would separate a qualification run from a programme.
Which chip would say whether EMIB is being tried on a mature part or the flagship. The node would say whether Intel is winning logic as well as assembly. The order size and the date would say whether Fab 9 is filling capacity or running a sample lot. Microsoft has committed to a supplier without committing to a volume.
The background explains part of Intel's urgency. The US government took a 10% stake in the company in August 2025, and Nvidia invested $5 billion in it. Intel needs named external packaging customers to justify the capacity at Fab 9. Microsoft needs a second source it can point to.
My reading is that this is a real validation of EMIB and a deliberately limited one. Microsoft has confirmed that Intel's packaging is good enough to put into production silicon, and has declined to say it is good enough to depend on.
The first public customer is a milestone, and milestones are cheap to announce. The second announcement, the one carrying a chip name and a date, is what will settle whether Intel's packaging business is a business or a demonstration.
*Disclaimer: This article is based on publicly available information and is for reference only. It does not constitute investment advice.*
