Meet Intel’s 18A-Based Starfire Orbital Chip Design

In space, no one can hear you churn out AI calculations, but you need the chips to do the job. Intel’s new Starfire is a pitch to be the company behind a new wave of potential orbital AI data centers, per company documentation spotted by a user on X. The chips are based on Intel’s new 18A process and support high operating temperatures and anti-radiation protection, but performance may be limited.

The big pitch for AI data centers in space from proponents like Elon Musk is that they can enjoy consistent solar power and don’t need to worry about noise levels, local pollution, or building permits. But they face other concerns instead: How do you cool data centers in space? How do you make them fast enough to compete with ground-based alternatives? And how do you make them even remotely cost-effective?

If building a data center on the ground takes tens of billions of dollars, building one in space, where every launch costs tens to hundreds of millions of dollars, seems unlikely to offer much return on investment. Not to mention the maintenance costs.

Still, if these projects do get off the ground, Intel wants xAI et al to use its chips. 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: 1GHz and 850 MHz for the LowPower version, and 3.1GHz and 2.1GHz for the Performance version.

Credit: Intel

They each have a three-tile NPU and three-tile GPU, clocked at 800MHz-1GHz and 2GHz, 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 of AI power for 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, apparently. Intel appended its slide on the new chips with “Domestic US manufacturing,” which is useful for potential buyers. It also feels targeted at a very specific audience.

All of this is impressive for a chip in space, but compared to anything on Earth? It’s pedestrian. A single Blackwell GPU can deliver over 1,000 TOPS on its own, and data centers include tens of thousands of them. These orbital projects might have a long-term use, but not until the hardware is vastly more powerful and efficient—and someone figures out how to cool them effectively.