Microsoft will use Intel Foundry's advanced packaging technology, EMIB, to package chips it designs for Azure. According to a joint announcement published by Intel on November 18, 2026, the two companies signed a long-term agreement covering part of Azure's infrastructure, securing multiyear Xeon supply that includes the next-generation "Diamond Rapids."
The collaboration also extends to Azure virtual machines (VMs) built on Xeon. A product that uses CXL to expand memory capacity has entered limited preview. The companies also published test results showing performance gains, but the VMs compared differ in memory configuration and CPU count, so the results need to be read with those conditions in mind.
Intel's EMIB to package Azure's custom-designed chips
EMIB connects multiple semiconductor dies side by side within a single package. According to Intel's technical explainer, small silicon bridges embedded in the package substrate link adjacent dies. According to Microsoft's own joint announcement, this is the first time it has disclosed the use of EMIB for Azure's custom-designed chips.
Microsoft and Intel previously announced a manufacturing collaboration in 2024. In Intel Foundry's announcement that February, Microsoft said it had selected a chip design to be manufactured on the Intel 18A process. While 18A is a front-end process that forms circuits on the wafer, EMIB is a back-end packaging technology that connects finished dies. The latest announcement clarifies which packaging technology will be used for Azure chips.
At the same Ignite 2026 event, Microsoft announced its custom inference chip, Maia 300, on November 17, and made Cobalt 200 VMs, which use its own Arm-based CPU, generally available. Azure is expanding its use of custom-designed chips while relying on outside companies' technology for manufacturing and packaging. For Intel Foundry, the deal opens a new business opportunity in assembling Microsoft's custom chips.
Multiyear Xeon supply secured, including Diamond Rapids
The multiyear agreement also covers instances powered by the next-generation Xeon, Diamond Rapids. Microsoft expects Cobalt 200 VMs to account for a substantial share of the compute capacity added to Azure starting in 2027. Xeon, meanwhile, remains an option for large-capacity VMs running existing x86 software, so Azure will keep a mix of Arm and x86.
According to Intel's Q3 2026 earnings prepared remarks, Diamond Rapids is at the sampling stage, with production set to ramp in 2027. It will be Intel's first server product built on the Intel 18A-P process. In a joint announcement with Oracle on October 27, Intel also outlined plans to use Diamond Rapids in Oracle Cloud Infrastructure's X14 and a multiyear collaboration.
As adoption plans take shape at both Azure and Oracle Cloud Infrastructure (OCI), Intel's ability to scale supply depends not only on wafer manufacturing capacity but also on back-end processes. In its earnings remarks, Intel said the advanced organic substrates used in EMIB are in short supply across the industry, constraining its ability to expand. To grow both Xeon supply and packaging for customer-designed chips, Intel will need to add packaging capacity and secure a stable substrate supply.
CXL-enabled VM performance up to 2.3x, but test conditions differ
Azure is offering "Mv4 with CXL," which supports Xeon 6 and CXL, in limited preview. CXL (Compute Express Link) is a standard for connecting CPUs to peripheral devices, and in this VM it is used to expand memory capacity. According to the Microsoft Mv4 technical blog, the VM supports up to 8TB of memory and 304 vCPUs, and is aimed mainly at large in-memory databases such as SAP HANA, as well as SQL Server and real-time analytics.
The "Intel Flat Memory Mode" used in this configuration treats DRAM attached directly to the CPU and CXL-attached memory as a single address space, with hardware managing data placement at cache-line granularity. CXL-attached memory has higher access latency than CPU-attached DRAM, but placing frequently used data appropriately can limit the impact. Research presented at OSDI 2024 has also explored ways to reduce the performance gap between memory tiers through hardware-managed data placement. As a result, the benefit of CXL memory expansion varies with an application's memory access patterns.
Both companies' November performance results feature a "2.3x" figure, but the tests differ. The up-to-2.3x figure in Intel's joint announcement comes from a sandbox environment test, while the 2.3x in Microsoft's technical blog comes from SQL processing on a 10TB data warehouse. The comparison conditions are as follows.
| Reported performance figure | VM configurations compared | Test conditions |
|---|---|---|
| Up to 2.3x (sandbox) | Mv4 with CXL and DRAM: 76 vCPUs, 1,764GB. DRAM-only Mv4: 64 vCPUs, 1,007GB | Each sandbox allocated 2 vCPUs and 24GB, using a resident Python runtime with a 12GB dataset. Testing completed November 10 |
| 2.3x (data warehouse) | CXL-enabled Mv4: 304 vCPUs. Previous-generation Mv3: 176 vCPUs | 10TB SQL processing similar to TPC-H |
In the sandbox test, the compared VMs differ in vCPU count and memory capacity; in the SQL test, they differ in VM generation and vCPU count as well. Both 2.3x figures therefore reflect whole-system performance differences, not the effect of CXL alone. For SQL processing, Microsoft says performance per licensed virtual core also improved by about 34%. For databases that are licensed by core count, total cost depends in part on whether the required memory capacity can be secured while keeping the number of licensed cores down.
Microsoft also reports that in a separate test with total memory capacity held constant and half of the DRAM replaced with CXL memory, the performance difference stayed within about 4%. In a test where CXL was used to increase memory capacity, query throughput rose by up to 35% and storage I/O fell by more than 90%. The former examines the performance impact of changing the memory configuration; the latter examines the benefit of adding memory capacity. In actual deployments, comparisons will need to account for required memory capacity and access characteristics, as well as vCPU count and licensing costs.
Gains in SQL backup speed and confidential VM security
According to the companies, SQL backup speeds on Azure VMs supporting Intel QuickAssist Technology (QAT) rose by up to 2.5x in testing. The test described in footnote 2 of the joint announcement, completed November 2, compares an environment with Xeon Gold 8362, 128 logical CPUs and 1,024GB against an Azure VM with Xeon 6973P-C, 384 processors and 2.06TB. Because this is a whole-system comparison across different CPUs and memory capacities, the multiplier cannot be read as the speedup from QAT alone.
In confidential computing, the companies are using Intel Trust Domain Extensions (TDX), which isolates and protects data in use. When a confidential VM moves between hosts, the system encrypts the VM's memory and CPU register state to preserve confidentiality. At the same Ignite 2026, Microsoft also demonstrated technology for attaching verified devices to confidential VMs. Using the TDISP standard, the effort extends the devices to which a confidential VM can offload network processing to verified Azure Boost devices.
The joint announcement does not disclose the contract value, purchase volumes or exact contract term. The product name, manufacturing process and shipping timeline of the Azure chips that will use EMIB are also undisclosed. No specifics were given on the availability timing or regions of Azure VMs with Diamond Rapids, or on pricing and general availability for the CXL-enabled Mv4, which remains in limited preview.
