First of all, I would like to thank @rklb_invest and @PhotonCap for sharing these insightful points.
Drawing from my experience working at AMD AECG until last year, let me add some technical commentary on this.
Once a satellite is launched, that's it. No one can go up and swap out a chip. But mission requirements keep changing. Communication protocols get updated, sensor algorithms improve, and unexpected situations arise. ASICs have their logic baked in, so no changes are possible after launch. CPUs are flexible but slow. FPGAs deliver hardware-level speed while allowing you to remotely reprogram the logic itself from orbit. It's the only chip you can fix in an environment where nothing can be fixed.
The company that made FPGAs better than anyone was Xilinx. Across space, defense, telecom, and data centers, Xilinx was the undisputed number one in FPGAs. AMD acquired the company for roughly $50 billion in 2022. For AMD, this meant securing an entirely new pillar, Programmable Logic, on top of CPUs and GPUs.
After the acquisition, the Xilinx organization was restructured into AMD AECG (Adaptive and Embedded Computing Group). AECG's flagship products are Adaptive SoCs. While legacy FPGAs only offered programmable logic, today's Adaptive SoCs integrate ARM CPUs, AI Engines, and Programmable Logic onto a single chip.
The flagship product line is the Versal series. For space applications, there's the XQR Versal, a radiation-tolerant version built on a 7nm process with an onboard AI inference engine.
In the context of Google Suncatcher, here's how it breaks down. The TPU handles AI computation, and the AMD EPYC serves as the host CPU. But inside a satellite system, there's a lot more going on simultaneously: sensor data preprocessing, inter-satellite laser communication protocol handling, and real-time correction of radiation-induced bit-flip errors. This is exactly where FPGAs and Adaptive SoCs come in. They process at hardware-level speed while allowing logic updates mid-mission. And since a single chip covers multiple functions, it drastically reduces SWaP (Size, Weight, and Power), the most critical constraint in satellite design.
Bottom line: $AMD holds two seats at the table in space data centers. EPYC as the host CPU, and Adaptive SoCs for system control and data processing. The fact that one company holds both of these positions is something neither Intel nor NVIDIA can easily replicate.
