Pentagon wants computers that work with almost no power or memory

The Pentagon’s most ambitious research arm wants to build computers that can think in the dark, operate on almost no power, and keep working even when their own hardware is failing, and it is now asking the technology world to help figure out how.

DARPA, the Defense Advanced Research Projects Agency, Pentagon’s research and technology arm, published a Request for Information on June 18, 2026, asking researchers, universities, companies, and individual inventors to submit ideas on what the agency calls low resource computing, a broad term covering any approach to running meaningful software on hardware that would normally be considered too small, too weak, or too unreliable to be useful. Responses are due by July 17, 2026, and DARPA plans to hold an invitation-only workshop in Hanover, New Hampshire, in August 2026 to discuss the most promising submissions in depth.

The problem DARPA is trying to solve sits at an unusual intersection of military necessity and technological irony. Modern warfare increasingly depends on computing power deployed at the tactical edge, meaning on soldiers, in missiles, inside autonomous systems, and aboard vehicles operating in remote locations where there is no access to power grids, no reliable communications link, and no opportunity to offload processing to a distant data center. At the same time, the commercial technology industry has spent the past decade racing in exactly the opposite direction, building ever-larger artificial intelligence systems that consume gigawatt-scale power, the equivalent of what a small city uses, and require massive climate-controlled data centers to operate. The gap between what commercial computing has become and what military field operations actually need has widened into something approaching a chasm.

DARPA’s RFI document frames this paradox with a striking historical comparison. The agency points to ENIAC, the Electronic Numerical Integrator and Computer, which in 1945 represented the cutting edge of human computational achievement while consuming approximately 150 kilowatts of electricity, weighing nearly 27,000 kilograms (30 tons), and performing only thousands of operations per second with less than one kilobyte of working memory. The document then jumps to the present and notes that a chip inside a musical greeting card, manufactured for pennies and running on a fraction of a milliwatt of power drawn from a coin-cell battery, already outperforms that 1945 machine by every measurable metric while being millions of times smaller. The computation itself has become nearly free. The problem is that nobody has seriously explored how much further that miniaturization and efficiency can go when pushed by deliberate design rather than commercial accident.

What DARPA is asking for is not incremental improvement to existing processors. The agency is explicit that submissions describing modest efficiency gains or standard reductions in size, weight, and power requirements are not of interest. What it wants are disruptive concepts, ideas that represent a fundamental shift in how computing gets done when resources are genuinely scarce rather than merely constrained by engineering convention. The RFI organizes its requirements into two broad categories, physical and logical, covering four specific challenge areas in each.

On the physical side, DARPA is looking for approaches to ultra-low-power computing, including systems that run on nanowatts of power, passively harvest energy from their environment rather than drawing from a battery, and operate completely independently of external power sources. It wants computing paradigms that function in kilobytes or even bytes of memory, a constraint that forces entirely new approaches to data structures and algorithms compared to what modern software engineering assumes. It is interested in methods for achieving reliable computation on hardware that is inherently noisy, degrading, or unreliable, which describes the condition of any electronic component that has survived extended exposure to the physical punishment of a combat environment. Perhaps most provocatively, DARPA asks for ideas about computing using low-precision manufacturing processes, legacy fabrication techniques, or what it calls primitive technological ecosystems, including specific examples such as extracting computation from a purely mechanical CD player, using biological antenna arrays, or applying computational origami to circuit design. Those examples are not decorative. They reflect a serious institutional interest in computing methods that do not depend on cutting-edge semiconductor fabrication, which matters enormously for supply chain resilience at a time when the United States is acutely conscious of its dependence on overseas chip manufacturing.

The logical resource categories are equally revealing about what DARPA anticipates the future battlefield will look like. The agency asks for computing frameworks that function when data sources cannot be trusted, which directly addresses the problem of operating in environments where an adversary may have corrupted or manipulated the information a system is receiving. It wants systems that accomplish complex tasks with minimal permissions or system access, computing that works under self-imposed restrictions rather than assuming administrative control over its own hardware. It asks for self-hosting architectures, systems capable of reprogramming and modifying themselves without relying on external tools or connections to a development environment, a capability that would allow a deployed system to adapt to new requirements in the field without any link back to the engineers who built it. And it asks for interfaces so simple that soldiers operating under extreme stress, with minimal training, can deploy and use these systems without complex setup or troubleshooting.

The RFI is managed by DARPA’s Multi X Office, a relatively new organizational unit within the agency that focuses on cross-domain research problems that do not fit neatly within any single technical discipline. The program manager listed as the technical point of contact is Dan Ridge. The office’s mandate to pursue multi-domain challenges makes it a natural home for low resource computing, which touches materials science, computer architecture, algorithm design, human factors engineering, and supply chain strategy simultaneously.

DARPA has a long record of turning market research exercises into programs that reshape entire technological fields. GPS began as a research project. The internet began as ARPANET, a DARPA-funded network. The question this RFI poses is whether the same agency can now push computing in the opposite direction from where the market has taken it, toward systems so efficient, so small, and so self-sufficient that they can operate where no data center could ever reach.