A look at Microsoft's Project Catapult, a multi-year effort to design FPGA chips in-house to power cloud services including Bing, Azure, Office 365, more
It was December 2012, and Doug Burger was standing in front of Steve Ballmer, trying to predict the future.
Context & Ripple Effects
This 2016 feature is the origin story of Microsoft's silicon program: Doug Burger convincing Steve Ballmer in December 2012 that Microsoft should design its own FPGA hardware, which became Project Catapult powering Bing, Azure, and Office 365. At the time, a software company building chips was heretical — the bet was that reconfigurable FPGAs could be reprogrammed in place as workloads changed.
The arc since then validates the bet: Catapult's FPGA fabric became the substrate for Project Brainwave, Microsoft's real-time AI platform, and by late 2023 the program had graduated from programmable off-the-shelf logic to fully custom designs with the Maia 100 AI accelerator and Cobalt server CPU.
First-order effects
- Bing, Azure, and Office 365 gain an in-house acceleration layer whose behavior Microsoft can change via firmware rather than waiting on processor vendors' roadmaps.
- Doug Burger's research group moves from experiment to production infrastructure, giving Microsoft a standing internal chip-design capability.
Second-order effects
- Catapult establishes the template Microsoft follows a decade later with the Maia 100 and Cobalt — custom silicon offered to Azure clients — shifting its cost structure and negotiating position against merchant chip suppliers.
- Rival hyperscalers face pressure to match vertically integrated hardware-software stacks, since FPGA-accelerated services can be retuned per workload faster than competitors running general-purpose fleets.
Third-order effects
- If the pattern holds, cloud economics consolidate around operators who own their silicon stack end-to-end — the difference between renting compute from a vendor and manufacturing your own margin.
- The FPGA-first approach also normalizes reconfigurable infrastructure as a stepping stone: hardware deployed for one era's workloads (search ranking) gets repurposed for the next (real-time AI inference) without replacement.
The trend: Hyperscale clouds are migrating from merchant processors to in-house, workload-specific silicon, and Project Catapult is the data point where that migration at Microsoft began.