Mercury Research published its Q1 2026 x86 numbers on Thursday and the headline is one we have been tracking for a while: AMD now ships roughly one in three server CPUs sold worldwide. The exact figure is 33.2 percent of units, up around six percentage points year over year, with Intel still holding the rest at about 66.8 percent. Sequentially, Intel was flat. The shift looks incremental on a single chart, but compounded across the past eight quarters it is the steadiest market share migration the server silicon market has seen since the original Opteron era.
The aggregate x86 picture is messier. Total shipments across desktops, laptops and servers fell more than 6 percent quarter over quarter, which Mercury called worse than the typical seasonal dip. Desktop client volumes were down nearly 20 percent year over year, and Intel reallocated capacity toward server parts during the late 2025 supply squeeze, which means the client side took the brunt of the pain. Server unit shipments, in contrast, were up more than 10 percent year over year. The split tells us the only thing keeping the x86 market healthy right now is the AI capex cycle. For anyone running a fleet refresh in 2026, that has a direct consequence: the parts you most want are also the parts where supply is tightest and price discipline is weakest.
Arm is the other story buried in the data. Mercury estimates Arm now accounts for 13.2 percent of all server CPU shipments, up from 12.5 percent in Q4 2025 and roughly double the share from a year ago. Mercury president Dean McCarron attributed almost all of that growth to Nvidia's Grace CPU, which ships inside the Blackwell NVL72 rack systems that hyperscalers and large AI labs are buying in volume. For a typical European enterprise that does not buy NVL72 racks directly, this still matters because the managed inference and training services we rent from AWS, Azure and Google Cloud are increasingly running on these Arm based platforms underneath. Workload portability between x86 and Arm is no longer an academic exercise, it is a procurement lever.
The cloud relevance is concrete. AWS Graviton has been our easiest example for years, but Azure Cobalt and Google Axion are both shipping in production regions now, and Nvidia Grace based instances are showing up in specialty AI tiers. If our container images, JVM tuning and observability agents do not all build and run cleanly on aarch64, we are leaving 20 to 40 percent unit economics on the table in steady state compute, and we are also locked out of the cheapest inference capacity when the GPU queues are full. For platform teams at large European retail and e-commerce operators, the practical to do is short: add aarch64 to the default CI matrix, validate the top 20 base images, and put an Arm SKU column into the next FinOps showback so engineering leaders see the delta in euros, not in slideware.
The memory warning from AMD is the part that should land hardest with infrastructure leadership. AMD explicitly flagged that server CPU shipments may decline in H2 2026 because of the memory supply crisis. DRAM and HBM allocations have been tightening all spring, with HBM essentially spoken for through 2027 by the largest AI customers. That has two knock on effects for us. First, OEM quotes for memory rich SKUs, the 1.5 TB and 3 TB per socket configurations that database and in memory analytics workloads need, are likely to keep climbing. Second, hyperscalers are not immune. Reserved instance pricing and committed use discounts on memory optimized families have historically been the last to move, but renewal cycles in late 2026 are the ones to watch closely. If we have any three year commits coming up for renewal between September and December, we should be modeling at least a 10 to 15 percent uplift in list price as the base case, and pulling forward any commits we are confident about while current pricing holds.
There is also a competitive read across to Intel. Intel held server share flat sequentially, which is better than losing more, but it did so by sacrificing client capacity. That is a defensible short term trade, however it signals that Granite Rapids and the next Xeon 6 refresh need to land cleanly to stop the bleeding. For us, dual sourcing remains the right default. We should not let any single vendor exceed roughly two thirds of new server purchases over a rolling four quarter window, and we should keep at least one Arm based reference workload in production so we have a credible third option when negotiations get tense.
The TL DR for our infrastructure planning forum next week: AMD is now a default option not an alternative, Arm is a real third platform we have to support in CI and FinOps, and memory pricing is the variable most likely to blow up our 2026 cloud budget. Build the slide deck around those three points and the rest of the architectural conversations get easier.
