Intel Starfire orbital processor targets space AI data centers with 18A silicon

Intel's Starfire orbital processor is a space-grade SoC family built on the 18A process, designed for orbital AI data centers. It offers two SKUs with up to 75 TOPS, radiation tolerance, and a 10+ year lifespan, but faces steep economic and engineering hurdles that make it a niche proposal rather than an imminent deployment.

What happened

Intel's Starfire is a pitch to supply chips for potential orbital AI data centers, per company documentation spotted by a user on X. The chips are based on Intel's new 18A process technology and support high operating temperatures and anti-radiation protection. Starfire comes in two flavors: Low Power and Performance. Each has a mix of four Performance cores and four LP Efficiency cores, but the clock speeds are vastly different: 1.0 GHz and 850 MHz for the Low Power version, and 3.1 GHz and 2.1 GHz for the Performance version. Each variant includes a three-tile NPU and a three-tile GPU, clocked at 800 MHz-1 GHz and 2 GHz respectively. This results in the Low Power chip delivering up to 45 TOPS of AI compute on a 10W TDP, with the Performance chip delivering 75 TOPS at 35W. The chips are rated to perform between -55 and 125 degrees Celsius and use LPDDR5 or DDR5 memory. They'll last 10 years or more, according to Intel. The slide also notes "Domestic US manufacturing." For comparison, a single Blackwell GPU can deliver over 1,000 TOPS on its own, underscoring the performance gap between space-grade and terrestrial AI hardware.

Why AI builders should care

For AI builders, Starfire illustrates how AI accelerators might evolve to tolerate extreme environments. The radiation-hardened design and wide temperature range (-55 to 125 C) could inform resilience requirements for edge deployments on Earth, such as in aerospace, defense, or remote industrial settings. The concept also highlights the tension between specialized space-grade AI hardware and the economics of space launches and cooling. If orbital AI data centers ever become viable, the hardware constraints will be very different from terrestrial clusters: lower power budgets, no convective cooling, and long maintenance cycles. Builders planning on-orbit compute for satellite constellations or space-based inference should watch how Intel addresses these challenges.

Practical implications

Starfire emphasizes the need for robust thermal design and radiation-hardened components when considering AI workloads outside Earth. The 4P-core 4E-core configuration with integrated NPU and GPU tiles shows Intel's approach to balancing general-purpose compute and AI acceleration in a constrained power envelope. The memory options (LPDDR5 or DDR5) suggest flexibility for different orbital mission profiles. However, the modest TOPS figures (45-75 TOPS) compared to terrestrial GPUs mean that any orbital AI data center would need many chips to match ground-based throughput, driving up launch mass and cost. Cooling in space remains an unsolved problem: without air, heat must be radiated away, which adds complexity and weight.

Caveats

The article frames Starfire as a niche proposal targeting a very specific audience (xAI and partners) rather than a near-term mass deployment. ROI is questionable given that every launch costs tens to hundreds of millions of dollars, and building a ground data center already costs tens of billions. No real-world orbital deployments are cited, and the cooling challenge for space data centers is not addressed by the chip design itself. The performance gap with terrestrial hardware is large: a single Blackwell GPU exceeds 1,000 TOPS, while Starfire's Performance variant maxes out at 75 TOPS. Until someone figures out how to cool and power large arrays of these chips in orbit, Starfire remains an interesting engineering exercise rather than a practical option for AI builders.