Inside Astra's Alameda office, an internal AI model trained on every company artifact, from budgets to requirements documents to technical specs to work instructions, is being used to rewrite engineering work at a pace the company says no human team can match. The 110 people still employed there have been told to write code or leave.
That is the operating model Chris Kemp, Astra's CEO, described in an on-record appearance at the 41st Space Symposium in April 2026 and elaborated in a recent SpaceNews feature on Astra's AI push. It is also the model the company's 2026 is being run against.
"I told our senior leadership they need to 10x their output in 2026," Kemp told SpaceNews. "Our cost structure is going to go to nothing." The statement frames the year ahead as a labor-replacement experiment rather than a tool-augmentation program. Most space companies still publicly describe AI as something that helps engineers work faster. Astra is positioning itself as one of the earliest high-stakes tests of a different thesis: that a small hardware company can be run almost entirely by AI agents and the engineers who build them.
Kemp's mandate has two interlocking parts. The first is to "eliminate all software," meaning replace enterprise SaaS with in-house, AI-built tooling. The second is that every employee is now expected to write and contribute software to the company stack, with termination as the consequence for noncompliance. "Every employee at Astra writes software, or they're fired," Kemp said, as quoted in the SpaceNews report.
The company has already shrunk toward that target. Headcount fell from a peak of roughly 400 to 110, with Kemp signaling further declines as AI and automated manufacturing take on more of the engineering workload. The internal AI model is trained on the company's accumulated documents and is being used to flag inconsistencies, augment engineering output, and, in some cases, replace it. The stated end-state, in Kemp's words: "Robots are putting things together and AI agents are doing all the work."
Astra is not starting from a position of strength on the balance sheet, but it is not starting from zero either. The company generated nearly $50 million in 2025 revenue from spacecraft electric thrusters and reported a "slight" adjusted EBITDA profit, according to SpaceNews. The thruster business is small, but it is real, and it is what Kemp is leaning on while he rebuilds the launch side of the company around the new model.
The launch bet is named and dated. Astra is targeting the first flight of Rocket 4 for the end of 2026 from Cape Canaveral, with the vehicle positioned as a mass-manufactured product made possible by the lower-cost, AI-driven organization. Rocket 3, the company's previous launcher, failed repeatedly before Astra was taken private in 2024 at less than 1% of its peak SPAC-era valuation, in a deal led by Kemp and co-founder Adam London. The end-of-year Rocket 4 target is therefore a Kemp forecast rather than a confirmed manifest item, and Astra's history of missed launch deadlines is the reason to weight it as such.
The "no purchased software" stance is the part most exposed to friction. "Eliminate all software" is a slogan that will be tested the first time a regulator, customer, or cyber incident shows up; many of the tools Astra plans to replace carry compliance, security, and audit obligations that in-house, AI-built replacements will have to meet from day one. The "fire anyone who won't code" rule is a real cultural and legal exposure in a company that still needs accountants, recruiters, technicians, and salespeople, even if those roles are smaller than they were. And the AI model is only as good as the artifacts engineers put into it. If the inputs are stale, contradictory, or simply missing, the model will be too.
Kemp's broader bet has at least one external validator. At Anthropic's developer conference in May 2026, CEO Dario Amodei said he expects 2026 to produce the first billion-dollar company run by a single human employee, with AI doing the rest of the work, as SpaceNews reported. That is contextual color rather than an endorsement of Astra's specific plan. It does, however, frame Astra as an early test of that thesis in a hardware-heavy industry, where the cost of a wrong bet is a delayed launch rather than a missed deck.
What would count as evidence the experiment is working is now a short, named list. Did Astra hit its end-of-2026 Rocket 4 launch from Cape Canaveral? Did headcount continue to fall without the thruster revenue falling with it? Did the "10x output" directive show up as a measurable jump in shipped engineering work, and did the company keep its adjusted EBITDA roughly flat through the transition? Did regulators or customers push back on the all-in-house software stack, and how did Astra respond? The answers, one way or another, will land within the next four quarters.