Sources: Microsoft to use chips from Xilinx for more than half of the co-processors used on Azure servers, which previously used Intel chips exclusively
Ian King / Bloomberg :
Context & Ripple Effects
This 2018 report is the opening move in Microsoft's long march off merchant silicon: Bloomberg sources say Xilinx programmable chips will take over more than half of the co-processor slots on Azure servers that Intel had supplied exclusively. At the time it read as a supplier swap; in hindsight it was the first crack in a single-vendor dependency.
The arc since then confirms the trajectory — reports of in-house chip designs for cloud servers and Surface PCs surfaced two years later, followed by work on an AI processor with AMD (the codenamed Athena effort), the Maia 100 and Cobalt announcements, and finally Microsoft turning to Intel's 18A process to manufacture a chip it designed itself. The Xilinx decision established the playbook: qualify alternatives first, then replace.
First-order effects
- Intel loses its exclusive grip on Azure's co-processor socket immediately, with Xilinx taking the majority share of those parts and gaining a marquee hyperscaler reference design.
- Microsoft gains negotiating leverage and workload-matched acceleration on Azure without waiting on Intel's roadmap for FPGA-class parts.
Second-order effects
- Intel's datacenter group faces proof that its largest cloud customers will dual-source rather than accept sole-supplier pricing, pressuring margins across its server business.
- Xilinx's win signals to other FPGA and accelerator vendors that hyperscaler sockets are contestable, encouraging rivals to court Amazon and Google with similar co-processor programs.
Third-order effects
- If the pattern holds, the endpoint is exactly what the later coverage shows: hyperscalers designing their own silicon (Maia 100, Cobalt) and reducing suppliers to contract manufacturers — a role Intel itself accepts by fabricating Microsoft's 18A-designed chip.
- Cloud infrastructure stratifies into commodity CPUs plus differentiated accelerators, shifting value in the server stack away from the general-purpose processor vendor toward whoever controls the workload-specific silicon.
The trend: Hyperscalers are converting supplier dependence into design authority — starting with second-sourcing co-processors, ending with in-house chips and vendors recast as foundries.