This is an excerpt from Shyam Sankar and Madeline Harts forthcoming book Mobilize: How the Reboot the American Industrial Base and Stop World War III. Mobilize is available to order now.
Everything about how Marine Colonel Drew Cukor ran , the Department of Defenses upstart AI initiative, put a target on his back. He infuriated the acquisition community, which is a powerful enemy in . Ultimately, the firestorm of criticism triggered a series of unfounded but unrelenting IG reports that would harry Cukor until his retirement. Some of the details that follow may seem obscure, but theyre essential to understanding the bureaucratic inertia and pettiness that hold our military back.
When Cukor launched Maven in 2017, the government still bought software like it bought hardware. This posed a problem. The phases of a hardware program are research, development, test, and evaluation (RDT&E), followed by production and sustainment. Costs are very high initially, and then they decline. The Department of Defense treated software the same way. It paid a lot up front for a systems integrator to build software, then it paid very little when the software went into production for patches and minor security upgrades. Software was treated as a static, finished product once it entered production.
Heres the problem: software (at least, good software) is not static. Its constantly improving, yet the cost is relatively flat across stages of development, which is why you pay a recurring subscription for commercial software instead of a large, upfront fee. This insight is the basis of the software-as-a-service model, and it enables constant improvement of the product. Development, testing, and production of software happen simultaneously, all the time. Understanding this, Cukor made the heretical argument to Congress that Maven should be procured as a continuously evolving capability, with a similar cost over its lifetime. Cukor procured software using Broad Agency Announcements (BAAs), a flexible contracting vehicle that categorized software as RDT&E. Although this categorization wasnt perfect, the BAA allowed the program costs to reflect how software was developed and deployed and allowed Cukor to make frequent changes to the product while it was in production.
Cukor would soon run into other problems with categorizing software as RDT&E. The departments general posture is that if the US government is paying for R&D, it should own the intellectual property (IP) that results from that work. The problem is that despite the categorization of the contract vehicle as R&D, Maven wasnt paying for commercial companies to perform R&D. When Palantir or Microsoft or Amazon showed up on day one of their work with Maven, they showed up with products that had decades and billions of dollars already invested. The R&D was already done. Yes, that product would get fine-tuned during the program and the companies would learn from the governments mission and data, but fundamentally, the government was paying for software, not R&D. To Cukor, the governments obsession with owning IP was an “overstated matter” more likely to harm the companies, and therefore national security, in the long term. As Cukor correctly notes, “If you [the company] cant monetize this after working with us, then whats the use of doing this? Why would you hand over your IP ever?”
To be clear, the companies did not own the governments data and were not free to, say, sell a terrorist-targeting algorithm to China. International Traffic in Arms Regulations (ITAR) were in place, and the governments interests were protected. But a company that built a deep learning algorithm maintained the IP to its proprietary model weights. For Palantir, this meant that we retained the IP to our core platform while giving the government rights to Maven-specific logic configured on top of it.
Safe to say, Cukors approach was correct. Almost a decade later, Maven remains the bes
