US soldiers who tried AI-enabled drones and command tools in a July desert exercise say the equipment overheated, flew less far than advertised and often was not connected to the Army's new network. The Wall Street Journal highlighted those problems, which it framed as a challenge to the Army's modernization push; the paper is paywalled, so this post leans on DefenseScoop, Army releases and other reporting for specifics. The pattern is a useful reality check for anyone who assumes AI-enabled systems perform in the field the way demos suggest.
We have covered the policy side of military AI in the Senate probe into Pentagon AI overreliance and an AI tool that nearly triggered a US boarding of a Chinese ship. This story is the hardware-and-workflow counterpart.
TL;DR: what is reported and what is not
| Question | Answer |
|---|---|
| What was the exercise? | Project Convergence Capstone 6 (PC-C6), Fort Irwin, California, July 20 to 29, 2026 |
| Who took part? | Nearly 10,000 troops from the US, Britain, Australia, Canada and New Zealand, per press reports; the 4th Infantry Division ran NGC2 |
| What went wrong? | Overheating, weight, short range, jamming and drones not on the network, per soldiers quoted by DefenseScoop |
| What is NGC2? | The Army's Next Generation Command and Control prototype |
| Are the drones fully autonomous? | Anduril says its Bolt-M requires a human in the loop |
| Official results? | Not released as of August 11, 2026 |
We could not read the Wall Street Journal article itself. The specifics below come from DefenseScoop, a DVIDS story by the Army's 5th Mobile Public Affairs Detachment, and a Hurriyet Daily News summary.
What the Army said it was testing
The Army's own account describes PC-C6 as its premier experimentation event of the year. Its DVIDS release, dated July 24, says the 4th Infantry Division employed the NGC2 architecture at division level under realistic conditions while NGC2 was assessed alongside more than 90 developing capabilities. Maj. Gen. Patrick Ellis described NGC2 as "a whole stack approach to command and control - transport, integration, data, and applications."
The release adds that the Army had recently announced an NGC2 common data layer meant to let it plug in commercial technology such as edge computing, third-party apps and AI models. That is where the AI sits: not as one robot brain, but as models consuming shared data to speed decisions.
Other reporting describes drones, autonomous vehicles, electronic warfare equipment, an ALTIUS-600 loitering drone searching for targets, an autonomous Canadian Cessna monitoring from above, and 20-ton driverless trucks from Rheinmetall.
How NGC2 got here: the Ivy Sting history
Capstone 6 was the biggest test yet of a system that had been built in smaller steps. A Business Insider report from November 2025, republished by Yahoo, explains that NGC2 is meant to replace command systems the Army called not "mobile, intuitive, or survivable enough" for future threats. Anduril and a team including Palantir won a contract worth nearly $100 million for it, and the 4th Infantry Division built a prototype at Fort Carson, Colorado.
The earlier Ivy Sting exercises showed the promise. In Ivy Sting 1 in September 2025, an M777 howitzer was connected to fire control within minutes of being placed, compared with over an hour before. In Ivy Sting 2, three artillery pieces were connected in about ten minutes or less using six command nodes. In those tests, reconnaissance drones flew 30 to 40 minutes per flight, feeds went to the cloud, and AI models helped speed target identification. Even then, officials flagged needs for airspace deconfliction, more data and computing power, and contested electromagnetic conditions, which were slated for later tests.
Read against that history, the Mojave complaints are not a surprise so much as the next layer of difficulty. Controlled events at a home station prove the concept; a 10-day test in a harsh desert with a free-acting opposing force, as described by Hurriyet, tests everything else. A British officer quoted there said heavy signal jamming helped identify the limits of new systems and how to operate without reliable communications, which is the point of running the test.
What soldiers reported
DefenseScoop's reporter spoke to 4th Infantry Division soldiers and found the same themes repeatedly.
- Heat. Overheating was a major constraint. Soldiers improvised cooling, including laying ice-dunked T-shirts over hardware. Hurriyet, citing The Telegraph, reports temperatures reached 50.5 degrees Celsius and that lithium-ion batteries struggled. Anduril told DefenseScoop it is accelerating thermal management updates.
- Weight. Drones added to already heavy loads, so soldiers left batteries and cases behind. The recommended load of Teledyne FLIR's Rogue 1 is about 30 pounds; the Bolt-M weighs 13 to 16.8 pounds plus a 4.4-pound ground station.
- Range and links. Wind, desert ridges, competing frequencies and simulated enemy jamming limited connections. Soldiers flew shorter distances than vendors advertise: Anduril lists the Bolt-M at over 12 miles, and one soldier flew it less than half that.
- Network. Anduril said the Bolt-M was connected to NGC2, yet none of the soldiers interviewed said their drones were on the network.
- Benefits. Soldiers also said drones let them see farther without revealing their positions, and one captain said his capabilities were "exponentially increasing."
Why heat and weight are AI problems too
It is tempting to file overheating under hardware, not AI. But edge AI is compute, and compute makes heat. Running vision models on a drone or ground station for target recognition draws power in a sealed enclosure in a desert, which is the worst thermal case. The same trade-off shows up in civilian local AI: sustained throughput on a laptop depends on cooling, as our guide to Nvidia's RTX Spark power envelopes explains.
Weight matters because a soldier carries every battery. A model that needs a bigger battery or a larger ground station is a tax on an infantryman's pack. Efficiency, not peak accuracy, is the figure of merit at the edge.
Server blocks with airflow lines representing thermal limits on edge compute in AI drones
Human in the loop, and the vendor-soldier gap
Anduril told DefenseScoop that the Bolt-M requires a human in the loop, and soldiers identified targets from video feeds, weighing whether an expensive one-shot drone was worth using against targets a rifle might handle. That matches the broader principle in our guide to human-in-the-loop AI: the more costly the action, the more a person should approve it.
DefenseScoop also notes friction between vendors and users. Teledyne called one soldier's claim about a 10-minute setup "suspect," and vendors were barred from the training area, which made it hard to tell operator error from equipment malfunction. That is a familiar problem in AI evaluation generally: when builders cannot watch use, bug reports blur into user complaints.
Gate with a waiting shape illustrating human in the loop approval before a military AI drone acts
What is not known
- The WSJ's full account. We did not read it; check it for any quotes or figures we have not repeated.
- Official NGC2 results. DefenseScoop says the Army has not released them.
- Failure rates. Soldier testimony is qualitative. There is no published count of overheated units.
- Budget figures. A search for a roughly four billion dollar request returned no source, so we are not repeating it.
Why this matters beyond the Army
The story illustrates three things relevant to anyone building with AI. First, field conditions break assumptions made in the lab, whether heat, connectivity or user workflow. Second, integration is harder than capability: a drone that works alone but is not on the network does not deliver the promised shared picture. Third, feedback loops need access; vendors who cannot observe use cannot easily fix it.
It also connects to the policy argument about how far military AI should be trusted. The courts have been busy with the Pentagon's relationship to AI vendors, as in the ruling that the Anthropic blacklist was illegal and the later D.C. Circuit decision. Drone threats also reach civilian infrastructure, as in the drone strikes on a Yandex data center, where AI infrastructure is the target rather than the tool.
What to do if you build or buy edge AI
- Test at the worst thermal and connectivity conditions before claiming range or uptime.
- Budget for power and weight in the model choice, not only accuracy.
- Keep a human approval step for any action that is costly or irreversible.
- Give users a channel to report failures and give builders enough access to reproduce them.
- If you deploy autonomous agents in security-sensitive settings, AgentBeam, the agent security platform from the explainx.ai team, is built to stop AI agents before they take dangerous actions.
Related reading
- Senate investigation into Pentagon AI overreliance
- AI tool nearly triggered a US boarding of a Chinese ship
- Human-in-the-loop AI: when to let the agent run
- Judge rules Pentagon blacklist of Anthropic illegal
- D.C. Circuit overturns the Anthropic Pentagon blacklist ruling
- Yandex Vladimir data center hit by third drone strike
- Nvidia RTX Spark power envelopes
Details reflect DefenseScoop, Army and press reporting as of October 12, 2026 and may change as official results are published.