Buying autonomy is easy to sell and hard to afford. Program offices, industry briefers, and congressional pitch decks talk about unit price, capability demonstrations, and attritable concepts. What rarely survives the marketing slide is the long tail of sustainment: recurring software upkeep, data pipelines, cybersecurity, contractor logistics, supply chain churn, and the human teams that must keep autonomous machines safe and mission capable. Those are the bills that arrive every month, year after year, and that can eclipse the original procurement price.

Start with the baseline. The Government Accountability Office has repeatedly shown that operating and support costs drive the lion’s share of a weapon system’s life cycle bill. Sustainment historically accounts for roughly 70 percent of total life cycle cost for major systems. Programs that focus only on flyaway cost or prototype price run the high risk of underfunding the persistent expense of keeping systems running.

Autonomy multiplies sustainment vectors rather than simplifying them. Traditional hardware sustainment problems, like spares and depot capacity, remain. On top of that come software-specific lines on the ledger: continuous integration and delivery pipelines, model retraining, data labeling, cloud and edge inference costs, security monitoring, and regulatory compliance. In plain terms, an autonomous ground robot or drone is not a box you buy and then park in a hangar. It is an integrated hardware and software service that requires continuous investment to stay useful and safe.

Concrete case study lessons exist. The F-35 program, often discussed for its hardware expense, also demonstrates how software-driven sustainment can balloon costs and complexity. DOD and watchdog reports tied a burdensome logistics IT system to significant sustainment headaches for F-35 maintainers, and sustainment estimates for the program run into the trillions over multi-decade life cycles. Those issues are a reminder that a sophisticated software backbone, if designed or contracted poorly, can become a recurring drain on readiness and budgets.

What are the hidden line items that procurement glosses over?

  • Software as recurring development. Modern autonomy is software first. That means continuous updates, bug fixes, and capability improvements instead of a one time transition to sustainment. The Army has explicitly moved to treat many software systems as continuously developed assets, because traditional sustainment funding rules choke modern software practices. In practice that model makes sustainment look like ongoing RDT&E long after fielding. Budgeting accordingly is necessary, but rare.

  • Data pipelines and labeling. Supervised perception stacks and many AI components need curated training data. Labeling sensor footage, cleaning logs, and building test sets is labor intensive. The initial training may be expensive, but periodic retraining to deal with data drift is a predictable recurring cost. Industry estimates and practitioner accounts indicate that data preparation often consumes the majority of project effort for operational AI.

  • Compute, cloud and bandwidth. Running perception models, fleets of agents, or federated learning at scale costs money. Real time inference at the edge, secure backhaul to cloud services, and redundancy for contested networks add infrastructure and bandwidth fees. Those costs scale with operational tempo. They are not well represented by a one time capital purchase.

  • Cybersecurity and certification. Autonomy raises new attack surfaces. Secure update paths, hardened MLOps, and continuous authority to operate activities demand both tooling and manpower. Remediation after a vulnerability is found can trigger unplanned costs and operational pauses. Government and independent audits are no longer optional line items.

  • Contractor logistics and perverse incentives. Many programs rely on long term contracting to supply software updates, spare parts, and field service reps. Poorly structured incentives or opaque performance metrics can mean the government pays for convenience, not resilience. Prior audits found instances where sustainment contracts and software backends produced costly workarounds and incentive payments that did not correspond to actual readiness gains.

  • Obsolescence and sensor refresh. Lidar, cameras, radios, and processors age, and procurement lead times for replacement parts can be long. Refreshing sensors to meet evolving threat or environmental conditions is an often overlooked recurring procurement cost. Designing for modularity and open interfaces helps, but modular architectures themselves can have short term cost premiums that must be justified against long term savings.

  • Human capital. Operators, maintainers, data engineers, and cyber teams are necessary to keep autonomy effective. Recruiting and retaining those skills is expensive and competes with private sector demand. High personnel churn drives training and knowledge transfer costs.

Why does this matter? Because autonomy programs are being sized and fielded on assumptions that downplay recurring fees. Policymakers and program managers will face choices: accept lower readiness rates, reprogram funds from other accounts, or raise the long term budget baseline. None of those are pleasant. The GAO and congressional committees have repeatedly flagged sustainment as the area where cost growth bites hardest.

What does sensible procurement look like?

  • Budget for sustainment from day one. Cost estimates must include software life cycle work, data refresh cycles, cloud and bandwidth costs, cybersecurity operations, and contractor logistics. Treat autonomy as a service, not a durable good.

  • Buy outcomes, not hours. Contracts should reward demonstrable readiness and resilience, not opaque invoiceable maintenance labor. Performance based logistics and measurable availability metrics aligned to truthfully reported data reduce perverse incentives.

  • Insist on modular open architectures. MOSA style design reduces lock in and enables cheaper sensor swaps and upgrades. The trade is front loaded complexity and planning, which is worth the investment if programs are meant to last decades.

  • Fund organic software sustainment and MLOps. The services need in house capacity to operate CI/CD, run secure model retraining, and verify autonomy updates. Otherwise programs are permanently dependent on outside contractors at increasing marginal cost.

  • Require realistic operational demonstrations. Prototype runs that skip logistics and sustainment assumptions are misleading. Full evaluation must include the O&S envelope, including recurring software and data costs.

Autonomy offers real operational benefits. It can reduce risk to humans and buy new tactical options. Those benefits come with non trivial sustainment obligations. If leaders want a future where machines do dangerous work, they must also accept a present where budgets pay for the quiet, recurring work that makes those machines useful. Ignoring the sustainment tab will not make it go away. It will only shove costs into readiness shortfalls, surprise supplemental requests, and political finger pointing when the bill finally arrives.