Imagine an AI at the Pentagon simultaneously issues two instructions: eliminate an advanced radar installation near Kaliningrad and reroute 20 Maersk container vessels around the Red Sea. <\/p>\n\n\n\n
Within four seconds, both orders receive approvals \u2014 a brigadier\u2019s click and a CFO\u2019s e-signature. Following that brief interval, a hypersonic missile lights up the Baltic sky and $200 million in freight changes course. <\/p>\n\n\n\n
This rapid decision-making is made possible because approximately $1,000 in cloud-GPU compute replaced analytical tasks previously costing millions in human hours.<\/p>\n\n\n\n
Now, war as a video game isn’t new: drone pilots in Iraq have guided strikes from distant bunkers. What’s transformative about the above scenario is how it demonstrates cognition as an affordable commodity \u2014 one that\u2019s only likely to get cheaper and better in the future. <\/p>\n\n\n\n
Today\u2019s AIs are like superhuman analysts<\/a>, and they give militaries the ability to run multiple parallel scenarios for any situation and choose the optimal path, quickly and cheaply. With them, the cost of answering a problem is reduced to computation cycles, and the time it takes to complete the OODA loop<\/a> \u2014 observe, orient, decide, and act \u2014 is cut down to seconds. <\/p>\n\n\n\n The number of AI analysts a military can have \u2014 each of which can do almost every task \u2014 will be directly proportional to the amount of compute it can get. Whoever leverages the low-cost capabilities of AI the most will have a decisive advantage. <\/p>\n\n\n\n We\u2019re already seeing concrete examples of this in action.<\/p>\n\n\n\n Palantir\u2019s Maven system demonstrated a sensor-to-shooter loop of roughly a dozen seconds in recent NATO acceptance trials<\/a>. XVIII Airborne Corps operators, during the 2025 Scarlet Dragon exercise<\/a>, achieved AI-assisted strikes approved in under 30 seconds, exponentially increasing decision throughput. Anduril\u2019s Lattice platform<\/a> targets engagement cycles of approximately 90 seconds from detection to strike.<\/p>\n\n\n\n AI is poised to do for military analytics what the steam engine did for logistics.<\/p>\n<\/blockquote>\n\n\n\n Still, as impressive as this is, knowledge amplifying military effectiveness is nothing new \u2014 it\u2019s been happening for as long as militaries have existed.<\/p>\n\n\n\n Knowing how to make gunpowder was once a huge military advantage as it rendered castles and armored knights obsolete, a fact best exemplified by the Ottoman bombards used at Constantinople in 1453. Napoleon leveraged operational strategy and mass mobilization to redefine warfare at Austerlitz. Industrial-era innovations, such as rail and telegraphy, restructured military logistics during the American Civil War, while World War II’s codebreaking and radar changed outcomes at Midway and Bletchley Park. <\/p>\n\n\n\n AI is just the next step-change in knowledge, and it is now poised to do for military analytics what the steam engine did for logistics: cut costs so dramatically it rewrites doctrine. Analytical prowess once costly and rare will become suddenly ubiquitous.<\/p>\n\n\n\n Affordable compute accelerates itself, continually widening strategic advantages.<\/p><\/blockquote><\/figure>\n\n\n\n As advanced as today\u2019s AI-based military solutions are, though, they\u2019re only going to get better and cheaper \u2014 and the pace<\/em> of improvement is likely to accelerate.<\/p>\n\n\n\n An example: The Department of Defense\u2019s Replicator program<\/a> is funding the development of AI-equipped air, sea, and land drones. Each flight streams telemetry from the drones into a common model-training pipeline. The data a $50K quadcopter collects today will refine the autonomy stack that designs its $30K successor next quarter. Data collected by that<\/em> drone will train its successor and so on\u2026\u00a0<\/p>\n\n\n\n We\u2019re seeing the same thing happen on the hardware side of AI development. <\/p>\n\n\n\n Affordable AI computation is accelerating chip design<\/a>. Those better chips are then used to lower computation costs further. Nvidia and Synopsys\u2019 AI-assisted Electronic Design Automation (EDA) tools have already shortened chip design cycles dramatically, perpetuating a cycle where affordable compute accelerates itself, continually widening strategic advantages. <\/p>\n\n\n\n For all the benefits AI offers the military, it brings with it new vulnerabilities. <\/p>\n\n\n\n DARPA red-team exercises have discussed concerns<\/a> about prompt-injection exploits \u2014 cyberattacks in which an LLM is prompted to do something nefarious purely through carefully constructed wording or manipulated images \u2014 with potential to override operator link and feed spoofed sensor data to redirect strike drones. <\/p>\n\n\n\n AI models capable of autonomous long-duration thought \u2014 ones that can engage in complex chains of reasoning without involving a human in the process \u2014 amplify these risks, turning unsupervised learning periods into significant strategic vulnerabilities.<\/p>\n\n\n\n Even the data used to train<\/em> an AI can be a point of weakness.<\/p>\n\n\n\n A 2022 study<\/a> found that malicious actors could \u201cpoison\u201d the data sets used to train an AI, slipping in a small amount of data that will eventually cause the model to engage in “sleeper-agent” behaviors. After replicating<\/a> this in larger models, AI startup Anthropic proposed continuous evaluations, through regular iterative, dynamic assessments, to detect manipulation early.<\/p>\n\n\n\n Ultimately, the solution seems to require knowing exactly what data goes into training an AI and out while asking it something. Ideally, this data should be timestamped, hashed, with a fingerprint that you can trace through the system. Palantir\u2019s hash-ledger, which is already in trial use<\/a>, might well become an essential defense.<\/p>\n\n\n\n\n
The challenges<\/h3>\n\n\n\n