
Intel.com
Microsoft and Intel announced an expanded long-term agreement on Wednesday, Nov. 18, in which Microsoft becomes the "first company to announce use of Intel Foundry's EMIB" advanced packaging for its own Azure silicon, according to Intel's joint press release. The release does not say which Azure chip will use Intel's packaging, which manufacturing node it will be built on, when it will ship or what Microsoft will pay. It also covers Intel Xeon supply, and it publishes two performance claims that compare unlike systems.
For Azure customers and anyone tracking custom cloud chips, the useful takeaway is a boundary: one named packaging commitment, plus Xeon supply that follows a separate path.
The announcement was published on Intel's newsroom on Nov. 18. Saurabh Dighe, corporate vice president of Azure System & Architecture at Microsoft, and Naga Chandrasekaran, Intel's executive vice president, chief technology and operations officer and general manager of Intel Foundry, are the executives quoted in the release.
The wording is narrow. Microsoft is the first to announce EMIB use for Azure first-party silicon, which is not the same as being the first to use it. Reports of Google and Amazon EMIB interest rest on industry sources, not announcements naming a product, and the release names no Azure chip or process node.
EMIB, short for Embedded Multi-die Interconnect Bridge, embeds small silicon bridges in the package substrate to connect chiplets. TSMC's CoWoS-S instead uses a silicon interposer to connect memory and compute dies. CoWoS-L, which TrendForce expects to remain mainstream for AI chips through 2028, embeds silicon bridge dies within the interposer, a design closer to EMIB's bridge approach.
The tradeoff is cost and size against maturity. Bernstein analysts estimate standard EMIB packaging costs in the low hundreds of dollars per chip, against roughly $900 to $1,000 for CoWoS on a Nvidia Rubin-class accelerator. TrendForce's package-size comparison put CoWoS-S at about 3.3 times a lithography reticle, against 6 times for EMIB-M today and a projected 8 to 12 times across 2026 and 2027. EMIB-T, the newer variant, adds through-silicon vias to the bridge for improved power delivery and signal integrity.
The constraint is yield. TrendForce reported on Sept. 23 that EMIB substrate yield reached about 45%, up from about 30% in the second quarter, with the supply chain targeting 60% by the first quarter of 2027. Those figures come from industry sources and refer to substrates, not finished packages. A later TrendForce report on Oct. 22 put EMIB-T substrate yields near 35% in the third quarter, short of a 50% year-end target. TrendForce also flagged a thermal-expansion mismatch between substrate material and silicon bridges as a manufacturing challenge. For a cloud operator, packaging yield decides how many accelerators ship per quarter, so the missing chip name and date matter more than the mechanism.
EMIB is a packaging service. Chiplet designs combine dies built on different process nodes in one package, as Diamond Rapids' mix of tiles shows, so Microsoft could use Intel's packaging without moving wafer production away from TSMC. That is inference; the release says nothing about Intel 18A or 14A wafers.
Existing public facts leave the question open. Microsoft's Maia 200 accelerator is built on TSMC's N3P process with 216 GB of HBM3e memory. In February 2024, Microsoft said it planned to make a chip it designed on Intel's 18A process, without naming the product. SemiAccurate has reported a future Maia part on 18A or 18A-P, which Microsoft has not confirmed. TrendForce, in September, expected both Microsoft and AWS to adopt CoWoS-L in 2027 and Maia 200 to stay on CoWoS-S.
Read together, the new agreement is most consistent with Microsoft adding a second packaging source for an unspecified chip. That is analysis, not a disclosed plan. No source reviewed ties the EMIB commitment to a specific Azure product.
The Xeon half of the agreement is a separate supply arrangement. Diamond Rapids, Intel's Xeon 7 generation, is due in 2027 on the 18A-P process with 256 cores. Intel uses UCIe-S, a copper link, to connect the fabric hub tiles to the core chiplets instead of EMIB, reportedly to get low-latency, uniform paths to every memory hub; that is a design goal, not a measured result. A Xeon order from Microsoft therefore says little about EMIB demand.
Microsoft is also not the first cloud provider tied to Diamond Rapids. Oracle said on Oct. 27 that planned OCI X14 offerings are expected to be based on the chip. On the other side of the x86 market, AMD said in a July 20 release that Azure HDv2 virtual machines will use its Venice processors, so the Intel supply deal does not make Azure an Intel-only fleet.
The release says CXL memory and Intel Flat Memory Mode lift Azure M-series v4 performance by up to 2.3 times, and that Intel QuickAssist Technology speeds SQL Server backups by up to 2.5 times, according to Intel's published footnotes. Both come with baselines that differ from the new configuration in more than the feature being credited.
Intel's first footnote pits an M76ds_2_v4 instance, with 76 virtual CPUs and 1,764 GB of memory, described as 1 TB of DRAM plus 1 TB of CXL memory, against an M64lds_v4 with 64 virtual CPUs and 1,007 GB of DRAM only. The workload was a sandbox microVM, and testing took place Nov. 10. The CXL instance has about 19% more virtual CPUs and about 75% more memory, so the result cannot isolate what CXL tiering contributes. Intel describes Flat Memory Mode as hardware-managed tiering between DRAM and CXL memory on Xeon 6. The footnote's own memory description, 1 TB plus 1 TB, does not sum to 1,764 GB.
The second footnote compares an Intel Xeon Gold 8362 system with 128 CPUs and 1,024 GB of memory against a Xeon 6973P-C system with 384 processors and about 2.06 TB, running SQL Server 2025, tested Nov. 2. That is three times the processors and about twice the memory on a newer generation, so the 2.5x figure does not credit QuickAssist alone. Both numbers are "up to" results from vendor-run tests, not independent benchmarks.
The release also names Intel Trust Domain Extensions, which encrypts VM memory and register state as confidential virtual machines move between hosts. Researchers have published attacks on it. A security assessment posted to arXiv found five vulnerabilities in TDX Module 1.5: one that lets a hypervisor fully compromise a Trust Domain and four that leak module memory.
A separate technique, DDRop, uses an interposer that costs under $200 to build and drops memory writes on a DDR5 bus. Researchers used it to gain control of a protected VM on TDX. The attack requires physical access to the server, so it mainly matters to customers whose threat model includes that.
The release confirms three things: Microsoft's EMIB announcement, multi-year Xeon supply and a joint effort on CXL, QuickAssist and TDX. It leaves four unknowns: which Azure chip uses EMIB, which foundry makes its wafers, when it ships and what it costs.
Customers weighing a move to CXL-equipped M-series v4 sizes should test their own workloads against a like-for-like baseline, because the published gains mix hardware changes. Readers tracking custom silicon should treat Google and Amazon EMIB reports as unconfirmed until those companies name a product. Microsoft and Intel can answer the open questions in one sentence naming the chip and the year.
EMIB embeds small silicon bridges in a package substrate, while CoWoS-S places chips on a full silicon interposer and CoWoS-L embeds bridge dies within the interposer. EMIB reportedly costs less per package, and CoWoS keeps the more mature supply chain.
No source reviewed says so. EMIB is a packaging process, and Microsoft's Maia 200 uses TSMC's N3P process; the release names no chip or process node.
Not for the link between its fabric hub tiles and core chiplets. Tom's Hardware reports that Intel uses UCIe-S there, and joins base tiles to core chiplets with Foveros Direct 3D; the processor is due in 2027.
Not directly. The 2.3x claim compares a 76-vCPU instance with 1,764 GB of memory, including CXL, against a 64-vCPU, 1,007 GB DRAM-only instance, and the 2.5x claim compares systems with three times the processors and about twice the memory.
