SpaceX Hands Debris Detection in Parts to Grok: 99.9% vs. 94% for Humans

When people talk about SpaceX, they usually think of rockets. But at the UP.Summit conference in Florida, the company's Vice President of Launch, Kiko Dontchev, said out loud what has long been apparent from its strategy: "Yes, we are a rocket company, but at our core we are an infrastructure company." What interests us more is something else: how exactly AI is used there, and what can be taken from it for your own projects.

What was actually automated

Formally, SpaceX acquired xAI in early 2026. Engineers were using xAI tools even before that, but according to Dontchev, the real leap came in the last few months. In his 16 years at the company he has never seen engineering work speed up like this: what used to take hours and days now takes seconds and minutes.

The most striking example is the search for FOD, that is, foreign objects inside parts. A tiny piece of metal or debris in a fuel line can kill an engine. Previously, an inspector would feed a borescope (a camera on a flexible probe) into a tube and look for defects on the video by eye. According to Dontchev, people find such debris in about 94% of cases. Grok, analyzing the same video, manages 99.9%.

The same video-based checks are now being automated for other hardware as well, for example for Starlink user terminals. Dontchev says AI is used at every level, all the way down to technicians and their managers, and that it has "leveled the playing field" between newcomers and experienced employees.

Why this is not just a nice-looking number

The main lesson here is methodological. The comparison is made on the same data: the human and the model watch the same video. This is exactly how AI quality control should be evaluated. You are not arguing "AI versus humans" in general; you take a specific task, measure the baseline human accuracy, and test the model on the same recordings.

At the same time, the figures should be read with caution. These are the words of an executive at a conference, not a report with a methodology. It is unknown how many videos were counted, and what exactly "99.9% perfect" means: the share of defects found or overall accuracy. Nothing is said about false positives either. In tasks like this they matter just as much: if the model shouts "debris!" on every second tube, production will grind to a halt. Nothing is known about whether a human remains in the loop for disputed cases.

Other context: Falcon 9 is on its way out, Starship is coming

In the same conversation, Dontchev explained why Falcon 9 is heading toward retirement: they never learned to recover the second stage, and it has to be built anew for every launch. For a hundred launches a year that is tolerable; for thousands, and in the dreams of some even 10,000, it is not. Starship is supposed to catch both stages.

It is telling that he spoke not about payload capacity but about "performance per launch." Starship's first orbital flight in late September delivered Starlink V3 satellites with a combined throughput 10 times greater than Falcon 9's. From Florida, by his account, it will be up to 20 times. Pad 39A at Cape Canaveral is almost ready, and the first stage shipments are planned for December.

There is also a worrying side. Ars Technica notes that Dontchev did not address the concerns of commercial customers: according to several sources, in the coming years Starship will be reserved for SpaceX's own needs and possibly for NASA and other government customers. The enthusiast community, for its part, worries that Musk is shifting from Mars to orbital data centers. Dontchev insists that Mars remains at the core of the plans and that AI will only speed up the path to it.

What to do with this

• If you have visual inspection (manufacturing, warehouse, media content, screenshot tests), start by measuring human accuracy on real recordings. Without a baseline, any "99%" means nothing.
• Compare the model and the human on the same data, and count missed defects and false alarms separately.
• Keep a human on disputed cases, especially where the cost of error is high.
• Treat bold figures from interviews as a hypothesis to be tested, not as a proven result.

The SpaceX story is worth reading as a signal: AI enters heavy industry not through chatbots but through boring, measurable tasks like inspecting tubes. That is exactly where the savings become verifiable.