SpaceX made a play for Cognition and the Devin team said no
Here's what makes this fascinating: SpaceX doesn't acqui-hire. They acquire technology they can vertically integrate into Starship, Starlink, and the Falcon pipeline. They bought Swarm for satellite comms, Pioneer Astronautics for life support, and a handful of niche materials outfits. Cognition doesn't build rockets. They build an agent that can spin up a repo, write tests, debug CI failures, and open a PR while you grab coffee.
So why the interest? Two words: ground software.
Every launch generates terabytes of telemetry. Every Starship iteration rewrites flight computers, guidance loops, and pad sequencing. The current toolchain is a patchwork of MATLAB, legacy C++, and custom Python scripts held together by tribal knowledge. Imagine pointing Devin at the Starship codebase: "Refactor the thrust vector control module for the new Raptor 3 spec, keep all existing tests passing, and don't touch the abort logic." That's not science fiction — that's exactly the workflow Cognition demoed last month on a real-world Django migration.
The Devin team turned it down. According to people close to the negotiations, the founders (Scott Wu, Steven Hao, and Walden Yan) believe they're building a general-purpose engineering agent, not a defense contractor's internal tool. They'd rather raise a Series B at a $2B valuation and sell seats to every software org on the planet than become SpaceX's internal IDE plugin.
Can't blame them. The TAM for "AI that writes production code" is essentially the entire global software spend — call it $1.5T annually. The TAM for "SpaceX's internal dev tools" is whatever Elon budgets next quarter.
What this signals is bigger than one failed acquisition. The serious players — not the wrapper startups, not the chatbot skins — are now acquisition targets for the world's most ambitious engineering organizations. Google didn't buy Character.ai for the personality; they bought the inference stack. Microsoft didn't partner with OpenAI for the brand; they needed the model weights. SpaceX trying to swallow Cognition whole tells you exactly where the leverage sits: the agent layer that turns intent into verified, deployed code.
If you're building in this space right now, the strategic question isn't "how do I add AI to my product." It's "which strategic acquirer needs my agent capability badly enough to pay a 50x revenue multiple?" Because that's the new floor.