Intel Unveils Starfire Chip Designed for AI Processing on US Military Satellites in Orbit

Intel has introduced a new system-on-chip called Starfire that is intended to operate not in data centers but in Earth orbit, where it must endure temperature swings from -55 °C to 125 °C and continuous exposure to cosmic radiation.

The chip was developed under contract for the US government and combines eight general-purpose CPU cores, a three-die neural-network accelerator, and a graphics block inside a single Foveros package. Two power variants are offered: a 10 W version rated at up to 45 TOPS and a 35 W version rated at up to 75 TOPS.

Manufacturing processes are deliberately split inside the same package. The CPU cores and neural accelerator are fabricated on Intel’s leading-edge Intel 18A node, while the graphics module uses the more mature Intel 3 process—the same heterogeneous approach Intel previously employed in the server processor Clearwater Forest. Because smaller transistors hold less charge, they are more susceptible to single-event upsets from space radiation; therefore, reliability will be achieved through circuit-level and architectural hardening rather than process maturity alone.

For decades the space-grade processor market has been dominated by BAE Systems’ RAD750, a 200 MHz device flying on Mars rovers and the Kepler and Fermi telescopes. Newer offerings such as the RAD5545 and a forthcoming Microchip design for NASA promise roughly an order-of-magnitude performance increase, yet they remain focused on telemetry and control. Starfire, with its dedicated neural accelerator, is aimed at a different mission: performing artificial-intelligence inference and real-time data analysis directly on board the satellite.

The device is still a demonstration vehicle. Radiation-tolerance figures are still being measured, and the chip has not yet received formal space-radiation certification. Intel cautions that final specifications may change. First engineering samples are slated for delivery in the third quarter of 2026 through the company’s Intel Government Technologies division.

The project carries strategic weight for Intel, which remains the only US company capable of producing chips of this complexity. It participates in the Pentagon’s RAMP-C and SHIP programs. Company roadmaps indicate that 18A process yields are expected to reach industry parity only in 2027, potentially affecting the timeline for volume production of space-qualified parts.